def main(): """ Visualization for comparison of anchor filtering with 2D vs 3D integral images Keys: F1: Toggle 3D integral image filtered anchors F2: Toggle 2D integral image filtered anchors F3: Toggle 2D integral image empty anchors """ anchor_2d_colour_scheme = {"Anchor": (0, 0, 255)} # Blue anchor_3d_colour_scheme = {"Anchor": (0, 255, 0)} # Green anchor_unfiltered_colour_scheme = {"Anchor": (255, 0, 255)} # Purple # Create Dataset dataset = DatasetBuilder.build_kitti_dataset( DatasetBuilder.KITTI_TRAINVAL) sample_name = "000001" img_idx = int(sample_name) print("Showing anchors for sample {}".format(sample_name)) # Options # These clusters are from the trainval set and give more 2D anchors than 3D clusters = np.array([[3.55, 1.835, 1.525], [4.173, 1.69, 1.49]]) anchor_stride = [3.0, 3.0] ground_plane = obj_utils.get_road_plane(img_idx, dataset.planes_dir) area_extents = np.array([[-40, 40], [-5, 3], [0, 70]]) anchor_3d_generator = grid_anchor_3d_generator.GridAnchor3dGenerator() # Generate anchors start_time = time.time() anchor_boxes_3d = anchor_3d_generator.generate(area_3d=area_extents, anchor_3d_sizes=clusters, anchor_stride=anchor_stride, ground_plane=ground_plane) end_time = time.time() print("Anchors generated in {} s".format(end_time - start_time)) # Get point cloud point_cloud = obj_utils.get_stereo_point_cloud(img_idx, dataset.calib_dir, dataset.disp_dir) ground_offset_dist = 0.2 offset_dist = 2.0 # Filter points within certain xyz range and offset from ground plane # Filter points within 0.2m of the road plane slice_filter = dataset.kitti_utils.create_slice_filter(point_cloud, area_extents, ground_plane, ground_offset_dist, offset_dist) points = np.array(point_cloud).T points = points[slice_filter] anchors = box_3d_encoder.box_3d_to_anchor(anchor_boxes_3d) # Create 2D voxel grid vx_grid_2d = voxel_grid_2d.VoxelGrid2D() vx_grid_2d.voxelize_2d(points, 0.1, area_extents) # Create 3D voxel grid vx_grid_3d = voxel_grid.VoxelGrid() vx_grid_3d.voxelize(points, 0.1, area_extents) # Filter the boxes here! start_time = time.time() empty_filter_2d = anchor_filter.get_empty_anchor_filter_2d( anchors=anchors, voxel_grid_2d=vx_grid_2d, density_threshold=1) anchors_2d = anchor_boxes_3d[empty_filter_2d] end_time = time.time() print("2D Anchors filtered in {} s".format(end_time - start_time)) print("Number of 2D anchors remaining: %d" % (anchors_2d.shape[0])) unfiltered_anchors_2d = anchor_boxes_3d[np.logical_not(empty_filter_2d)] # 3D filtering start_time = time.time() empty_filter_3d = anchor_filter.get_empty_anchor_filter( anchors=anchors, voxel_grid_3d=vx_grid_3d, density_threshold=1) anchor_boxes_3d = anchor_boxes_3d[empty_filter_3d] end_time = time.time() print("3D Anchors filtered in {} s".format(end_time - start_time)) print("Number of 3D anchors remaining: %d" % (anchor_boxes_3d.shape[0])) anchor_2d_objects = [] for anchor_idx in range(len(anchors_2d)): anchor = anchors_2d[anchor_idx] obj_label = box_3d_encoder.box_3d_to_object_label(anchor, 'Anchor') # Append to a list for visualization in VTK later anchor_2d_objects.append(obj_label) anchor_3d_objects = [] for anchor_idx in range(len(anchor_boxes_3d)): anchor = anchor_boxes_3d[anchor_idx] obj_label = box_3d_encoder.box_3d_to_object_label(anchor, 'Anchor') # Append to a list for visualization in VTK later anchor_3d_objects.append(obj_label) unfiltered_anchor_objects = [] for anchor_idx in range(len(unfiltered_anchors_2d)): anchor = unfiltered_anchors_2d[anchor_idx] obj_label = box_3d_encoder.box_3d_to_object_label(anchor, 'Anchor') # Append to a list for visualization in VTK later unfiltered_anchor_objects.append(obj_label) # Create VtkAxes axes = vtk.vtkAxesActor() axes.SetTotalLength(5, 5, 5) # Create VtkBoxes for boxes vtk_2d_boxes = VtkBoxes() vtk_2d_boxes.set_objects(anchor_2d_objects, anchor_2d_colour_scheme) vtk_3d_boxes = VtkBoxes() vtk_3d_boxes.set_objects(anchor_3d_objects, anchor_3d_colour_scheme) vtk_unfiltered_boxes = VtkBoxes() vtk_unfiltered_boxes.set_objects(unfiltered_anchor_objects, anchor_unfiltered_colour_scheme) vtk_voxel_grid = VtkVoxelGrid() vtk_voxel_grid.set_voxels(vx_grid_3d) vtk_voxel_grid_2d = VtkVoxelGrid() vtk_voxel_grid_2d.set_voxels(vx_grid_2d) # Create Voxel Grid Renderer in bottom half vtk_renderer = vtk.vtkRenderer() vtk_renderer.AddActor(vtk_2d_boxes.vtk_actor) vtk_renderer.AddActor(vtk_3d_boxes.vtk_actor) vtk_renderer.AddActor(vtk_unfiltered_boxes.vtk_actor) vtk_renderer.AddActor(vtk_voxel_grid.vtk_actor) vtk_renderer.AddActor(vtk_voxel_grid_2d.vtk_actor) vtk_renderer.AddActor(axes) vtk_renderer.SetBackground(0.2, 0.3, 0.4) # Setup Camera current_cam = vtk_renderer.GetActiveCamera() current_cam.Pitch(170.0) current_cam.Roll(180.0) # Zooms out to fit all points on screen vtk_renderer.ResetCamera() # Zoom in slightly current_cam.Zoom(2.5) # Reset the clipping range to show all points vtk_renderer.ResetCameraClippingRange() # Setup Render Window vtk_render_window = vtk.vtkRenderWindow() vtk_render_window.SetWindowName("Anchors") vtk_render_window.SetSize(900, 500) vtk_render_window.AddRenderer(vtk_renderer) # Setup custom interactor style, which handles mouse and key events vtk_render_window_interactor = vtk.vtkRenderWindowInteractor() vtk_render_window_interactor.SetRenderWindow(vtk_render_window) vtk_render_window_interactor.SetInteractorStyle( vis_utils.ToggleActorsInteractorStyle([ vtk_2d_boxes.vtk_actor, vtk_3d_boxes.vtk_actor, vtk_unfiltered_boxes.vtk_actor, ])) # Render in VTK vtk_render_window.Render() vtk_render_window_interactor.Start()
def load_samples(self, indices): """ Loads input-output data for a set of samples. Should only be called when a particular sample dict is required. Otherwise, samples should be provided by the next_batch function Args: indices: A list of sample indices from the dataset.sample_list to be loaded Return: samples: a list of data sample dicts """ sample_dicts = [] for sample_idx in indices: sample = self.sample_list[sample_idx] sample_name = sample.name # Only read labels if they exist if self.has_labels: # Read mini batch first to see if it is empty anchors_info = self.get_anchors_info(sample_name) if (not anchors_info) and self.train_val_test == 'train' \ and (not self.train_on_all_samples): empty_sample_dict = { constants.KEY_SAMPLE_NAME: sample_name, constants.KEY_ANCHORS_INFO: anchors_info } return [empty_sample_dict] obj_labels = obj_utils.read_labels(self.label_dir, int(sample_name)) # Only use objects that match dataset classes obj_labels = self.kitti_utils.filter_labels(obj_labels) else: obj_labels = None anchors_info = [] label_anchors = np.zeros((1, 6)) label_boxes_3d = np.zeros((1, 7)) label_boxes_2d = np.zeros((1, 4)) label_classes = np.zeros(1) img_idx = int(sample_name) lidar_only = False num_views = 1 if not lidar_only: # Load image (BGR -> RGB) cv_bgr_image = cv2.imread(self.get_rgb_image_path(sample_name)) rgb_image = cv_bgr_image[..., ::-1] image_shape = rgb_image.shape[0:2] # Append the depth channel if self.add_depth: depth_map = obj_utils.get_depth_map( img_idx, self.depth_dir) # Set invalid pixels to max depth depth_map[np.asarray(depth_map == 0.0)] = \ self.kitti_utils.bev_extents[1, 1] # Add channel dimension to make stacking easier depth_map = np.expand_dims(depth_map, 2) image_input = np.concatenate([rgb_image, depth_map], axis=2) else: image_input = rgb_image else: image_shape = (370, 1224) # Get ground plane ground_plane = obj_utils.get_road_plane(int(sample_name), self.planes_dir) #ground_plane = np.array([0,-1,0,1.68]) if lidar_only: p_matrix = np.zeros((num_views, 3, 4), dtype=float) if num_views > 0: p_matrix[0] = np.array([[ 8.39713500e+02, 3.58853400e+01, 4.48566750e+02, 2.31460650e+03 ], [ 1.02835238e-13, 8.54979440e+02, 1.57320433e+02, 2.49655872e+03 ], [ 0.00000000e+00, 7.97452000e-02, 9.96815000e-01, 5.14357000e+00 ]]) p_matrix[1] = np.array([[ 1.20171708e+03, 9.73326000e+01, 3.99933320e+02, 1.04945816e+04 ], [ 1.41054657e+01, 8.65088160e+02, 8.46334690e+01, 5.24229862e+03 ], [ 1.62221000e-01, 1.62221000e-01, 9.73329000e-01, 1.13555000e+01 ]]) else: # Get calibration stereo_calib_p2 = calib_utils.read_calibration( self.calib_dir, int(sample_name)).p2 point_cloud = self.kitti_utils.get_point_cloud( self.bev_source, img_idx, image_shape) # Augmentation (Flipping) if kitti_aug.AUG_FLIPPING in sample.augs: if not lidar_only: image_input = kitti_aug.flip_image(image_input) point_cloud = kitti_aug.flip_point_cloud(point_cloud) obj_labels = [ kitti_aug.flip_label(obj, image_shape) for obj in obj_labels ] ground_plane = kitti_aug.flip_ground_plane(ground_plane) if lidar_only: for i in range(num_views): p_matrix[i] = kitti_aug.flip_stereo_calib_p2( p_matrix[i], image_shape) else: stereo_calib_p2 = kitti_aug.flip_stereo_calib_p2( stereo_calib_p2, image_shape) # Augmentation (Image Jitter) if (kitti_aug.AUG_PCA_JITTER in sample.augs) and not lidar_only: image_input[:, :, 0:3] = kitti_aug.apply_pca_jitter( image_input[:, :, 0:3], aug_img_noise=self.aug_img_noise) # Augmentation (Random Occlusion) if kitti_aug.AUG_RANDOM_OCC in sample.augs: point_cloud = kitti_aug.occ_aug(point_cloud, stereo_calib_p2, obj_labels) if obj_labels is not None: label_boxes_3d = np.asarray([ box_3d_encoder.object_label_to_box_3d(obj_label) for obj_label in obj_labels ]) label_boxes_2d = np.asarray([ box_3d_encoder.object_label_to_box_2d(obj_label) for obj_label in obj_labels ]) label_classes = [ self.kitti_utils.class_str_to_index(obj_label.type) for obj_label in obj_labels ] label_classes = np.asarray(label_classes, dtype=np.int32) # Return empty anchors_info if no ground truth after filtering if len(label_boxes_3d) == 0: anchors_info = [] if self.train_on_all_samples: # If training without any positive labels, we cannot # set these to zeros, because later on the offset calc # uses log on these anchors. So setting any arbitrary # number here that does not break the offset calculation # should work, since the negative samples won't be # regressed in any case. dummy_anchors = [[-1000, -1000, -1000, 1, 1, 1]] label_anchors = np.asarray(dummy_anchors) dummy_boxes = [[-1000, -1000, -1000, 1, 1, 1, 0]] label_boxes_3d = np.asarray(dummy_boxes) label_boxes_2d = np.asarray([[-1.0, -1.0, -1.0, -1.0]]) else: label_anchors = np.zeros((1, 6)) label_boxes_3d = np.zeros((1, 7)) label_boxes_2d = np.zeros((1, 4)) label_classes = np.zeros(1) else: label_anchors = box_3d_encoder.box_3d_to_anchor( label_boxes_3d, ortho_rotate=True) # Create BEV maps bev_images = self.kitti_utils.create_bev_maps( point_cloud, ground_plane) height_maps = bev_images.get('height_maps') #bev random masking """ bev_drop_p = 0.5 rand_01 = random.random() mask_bev_layer = np.zeros(height_maps[0].shape,dtype=np.float32) if rand_01 > bev_drop_p: mask_idx = random.randint(0,4) height_maps[mask_idx] = mask_bev_layer """ #print(height_maps[0].shape) density_map = bev_images.get('density_map') bev_input = np.dstack((*height_maps, density_map)) #bev_input = np.transpose(np.array(height_maps),(1,2,0)) point_cloud = self.kitti_utils._apply_slice_filter( point_cloud, ground_plane).T if lidar_only: depth_map = np.zeros( (num_views, image_shape[0], image_shape[1]), dtype=float) for i in range(num_views): depth_map[i, :, :] = project_depths( point_cloud, p_matrix[i], image_shape[0:2]) depth_map_expand_dims = np.expand_dims(depth_map, axis=-1) sample_dict = { constants.KEY_LABEL_BOXES_3D: label_boxes_3d, constants.KEY_LABEL_ANCHORS: label_anchors, constants.KEY_LABEL_CLASSES: label_classes, constants.KEY_IMAGE_INPUT: depth_map_expand_dims, constants.KEY_BEV_INPUT: bev_input, constants.KEY_ANCHORS_INFO: anchors_info, constants.KEY_POINT_CLOUD: point_cloud, constants.KEY_GROUND_PLANE: ground_plane, constants.KEY_STEREO_CALIB_P2: p_matrix[0:num_views], constants.KEY_SAMPLE_NAME: sample_name, constants.KEY_SAMPLE_AUGS: sample.augs, constants.KEY_DPT_INPUT: depth_map } else: depth_map = project_depths(point_cloud, stereo_calib_p2, image_shape[0:2]) depth_map = np.expand_dims(depth_map, axis=0) sample_dict = { constants.KEY_LABEL_BOXES_3D: label_boxes_3d, constants.KEY_LABEL_BOXES_2D: label_boxes_2d, constants.KEY_LABEL_ANCHORS: label_anchors, constants.KEY_LABEL_CLASSES: label_classes, constants.KEY_IMAGE_INPUT: image_input, constants.KEY_BEV_INPUT: bev_input, constants.KEY_ANCHORS_INFO: anchors_info, constants.KEY_POINT_CLOUD: point_cloud, constants.KEY_GROUND_PLANE: ground_plane, constants.KEY_STEREO_CALIB_P2: stereo_calib_p2, constants.KEY_SAMPLE_NAME: sample_name, constants.KEY_SAMPLE_AUGS: sample.augs, constants.KEY_DPT_INPUT: depth_map } sample_dicts.append(sample_dict) return sample_dicts
def test_get_empty_anchor_filter_in_2d(self): # create generic ground plane (normal vector is straight up) area_extent = [(0., 2.), (-1., 0.), (0., 2.)] # Creates a voxel grid in following format at y = bin (-1.5, -0.5] # [ ][ ][ ][ ] # [ ][ ][x][ ] # [ ][ ][ ][ ] # [ ][ ][x][ ] pts = np.array([[0.51, -0.5, 1.1], [1.51, -0.5, 1.1]]) voxel_size = 0.5 voxel_grid = VoxelGrid() voxel_grid.voxelize(pts, voxel_size, extents=area_extent) # Define anchors to test boxes_3d = np.array([ [0.51, 0, 0.51, 1, 1, 1, 0], [0.51, 0, 0.51, 1, 1, 1, np.pi / 2.], [0.51, 0, 1.1, 1, 1, 1, 0], [0.51, 0, 1.1, 1, 1, 1, np.pi / 2.], [1.51, 0, 0.51, 1, 1, 1, 0], [1.51, 0, 0.51, 1, 1, 1, np.pi / 2.], [1.51, 0, 1.1, 1, 1, 1, 0], [1.51, 0, 1.1, 1, 1, 1, np.pi / 2.], ]) anchors = box_3d_encoder.box_3d_to_anchor(boxes_3d) # test anchor locations, number indicates the anchors indices # [ ][ ][ ][ ] # [ ][1][3][ ] # [ ][ ][ ][ ] # [ ][5][7][ ] gen_filter = anchor_filter.get_empty_anchor_filter(anchors, voxel_grid, density_threshold=1) expected_filter = np.array( [False, False, True, True, False, False, True, True]) self.assertTrue((gen_filter == expected_filter).all()) boxes_3d = np.array([ [0.5, 0, 0.5, 2, 1, 1, 0], # case 1 [0.5, 0, 0.5, 2, 1, 1, np.pi / 2.], [0.5, 0, 1.5, 1, 2, 1, 0], # case 2 [0.5, 0, 1.5, 1, 2, 1, np.pi / 2.], [1.5, 0, 0.5, 2, 1, 1, 0], # case 3 [1.5, 0, 0.5, 2, 1, 1, np.pi / 2.], [1.5, 0, 1.5, 1, 2, 1, 0], # case 4 [1.5, 0, 1.5, 1, 2, 1, np.pi / 2.] ]) anchors = box_3d_encoder.box_3d_to_anchor(boxes_3d) # case 1 # [ ][ ][ ][ ] [ ][ ][ ][ ] # [ ][o][ ][ ] [ ][o][o][ ] # [ ][o][ ][ ] [ ][ ][ ][ ] # [ ][ ][ ][ ] [ ][ ][ ][ ] # case 2 # [ ][ ][ ][ ] [ ][ ][ ][ ] # [ ][ ][o][o] [ ][ ][o][ ] # [ ][ ][ ][ ] [ ][ ][o][ ] # [ ][ ][ ][ ] [ ][ ][ ][ ] # case 3 # [ ][ ][ ][ ] [ ][ ][ ][ ] # [ ][ ][ ][ ] [ ][ ][ ][ ] # [ ][o][ ][ ] [ ][o][o][ ] # [ ][o][ ][ ] [ ][ ][ ][ ] # case 4 # [ ][ ][ ][ ] [ ][ ][ ][ ] # [ ][ ][ ][ ] [ ][ ][ ][ ] # [ ][ ][o][o] [ ][ ][o][ ] # [ ][ ][ ][ ] [ ][ ][o][ ] gen_filter = anchor_filter.get_empty_anchor_filter(anchors, voxel_grid, density_threshold=1) expected_filter = np.array( [False, True, True, True, False, True, True, True]) self.assertTrue((gen_filter == expected_filter).all())
def np_box_3d_to_box_8c(box_3d): """Computes the 3D bounding box corner positions from box_3d format. This function does not preserve corners order but rather the corners are rotated to the nearest 90 degree angle. This helps in calculating the closest corner to corner when comparing the corners to the ground- truth boxes. Args: box_3d: ndarray of size (7,) representing box_3d in the format [x, y, z, l, w, h, ry] Returns: corners_3d: An ndarray or a tensor of shape (3 x 8) representing the box as corners in following format -> [[x1,...,x8],[y1...,y8], [z1,...,z8]]. """ format_checker.check_box_3d_format(box_3d) # This function is vectorized and returns an ndarray anchor = box_3d_encoder.box_3d_to_anchor(box_3d, ortho_rotate=True)[0] centroid_x = anchor[0] centroid_y = anchor[1] centroid_z = anchor[2] dim_x = anchor[3] dim_y = anchor[4] dim_z = anchor[5] half_dim_x = dim_x / 2 half_dim_z = dim_z / 2 # 3D BB corners x_corners = np.array([ half_dim_x, half_dim_x, -half_dim_x, -half_dim_x, half_dim_x, half_dim_x, -half_dim_x, -half_dim_x ]) y_corners = np.array([0.0, 0.0, 0.0, 0.0, -dim_y, -dim_y, -dim_y, -dim_y]) z_corners = np.array([ half_dim_z, -half_dim_z, -half_dim_z, half_dim_z, half_dim_z, -half_dim_z, -half_dim_z, half_dim_z ]) ry = box_3d[6] # Find nearest 90 degree half_pi = np.pi / 2 ortho_ry = np.round(ry / half_pi) * half_pi # Find rotation to make the box ortho aligned ry_diff = ry - ortho_ry # Compute transform matrix # This includes rotation and translation rot = np.array([[np.cos(ry_diff), 0, np.sin(ry_diff), centroid_x], [0, 1, 0, centroid_y], [-np.sin(ry_diff), 0, np.cos(ry_diff), centroid_z]]) # Create a ones column ones_col = np.ones(x_corners.shape) # Append the column of ones to be able to multiply box_8c = np.dot(rot, np.array([x_corners, y_corners, z_corners, ones_col])) # Ignore the fourth column box_8c = box_8c[0:3] return box_8c
def main(): """ Visualization of anchor filtering using 3D integral images """ anchor_colour_scheme = { "Car": (0, 255, 0), # Green "Pedestrian": (255, 150, 50), # Orange "Cyclist": (150, 50, 100), # Purple "DontCare": (255, 0, 0), # Red "Anchor": (0, 0, 255), # Blue } # Create Dataset dataset = DatasetBuilder.build_kitti_dataset(DatasetBuilder.KITTI_TRAINVAL) # Options clusters, _ = dataset.get_cluster_info() sample_name = "000000" img_idx = int(sample_name) anchor_stride = [0.5, 0.5] ground_plane = obj_utils.get_road_plane(img_idx, dataset.planes_dir) anchor_3d_generator = grid_anchor_3d_generator.GridAnchor3dGenerator( anchor_3d_sizes=clusters, anchor_stride=anchor_stride) area_extents = np.array([[-40, 40], [-5, 3], [0, 70]]) # Generate anchors in box_3d format start_time = time.time() anchor_boxes_3d = anchor_3d_generator.generate(area_3d=area_extents, ground_plane=ground_plane) end_time = time.time() print("Anchors generated in {} s".format(end_time - start_time)) point_cloud = obj_utils.get_lidar_point_cloud(img_idx, dataset.calib_dir, dataset.velo_dir) offset_dist = 2.0 # Filter points within certain xyz range and offset from ground plane offset_filter = obj_utils.get_point_filter(point_cloud, area_extents, ground_plane, offset_dist) # Filter points within 0.2m of the road plane road_filter = obj_utils.get_point_filter(point_cloud, area_extents, ground_plane, 0.1) slice_filter = np.logical_xor(offset_filter, road_filter) point_cloud = point_cloud.T[slice_filter] # Generate Voxel Grid vx_grid_3d = voxel_grid.VoxelGrid() vx_grid_3d.voxelize(point_cloud, 0.1, area_extents) # Anchors in anchor format all_anchors = box_3d_encoder.box_3d_to_anchor(anchor_boxes_3d) # Filter the boxes here! start_time = time.time() empty_filter = \ anchor_filter.get_empty_anchor_filter(anchors=all_anchors, voxel_grid_3d=vx_grid_3d, density_threshold=1) anchor_boxes_3d = anchor_boxes_3d[empty_filter] end_time = time.time() print("Anchors filtered in {} s".format(end_time - start_time)) # Visualize GT boxes # Grab ground truth ground_truth_list = obj_utils.read_labels(dataset.label_dir, img_idx) # ---------- # Test Sample extraction # Visualize from here vis_utils.visualization(dataset.rgb_image_dir, img_idx) plt.show(block=False) image_path = dataset.get_rgb_image_path(sample_name) image_shape = np.array(Image.open(image_path)).shape rgb_boxes, rgb_normalized_boxes = \ anchor_projector.project_to_image_space(all_anchors, dataset, image_shape, img_idx) # Overlay boxes on images anchor_objects = [] for anchor_idx in range(len(anchor_boxes_3d)): anchor_box_3d = anchor_boxes_3d[anchor_idx] obj_label = box_3d_encoder.box_3d_to_object_label( anchor_box_3d, 'Anchor') # Append to a list for visualization in VTK later anchor_objects.append(obj_label) for idx in range(len(ground_truth_list)): ground_truth_obj = ground_truth_list[idx] # Append to a list for visualization in VTK later anchor_objects.append(ground_truth_obj) # Create VtkAxes axes = vtk.vtkAxesActor() axes.SetTotalLength(5, 5, 5) # Create VtkBoxes for boxes vtk_boxes = VtkBoxes() vtk_boxes.set_objects(anchor_objects, anchor_colour_scheme) vtk_point_cloud = VtkPointCloud() vtk_point_cloud.set_points(point_cloud) vtk_voxel_grid = VtkVoxelGrid() vtk_voxel_grid.set_voxels(vx_grid_3d) # Create Voxel Grid Renderer in bottom half vtk_renderer = vtk.vtkRenderer() vtk_renderer.AddActor(vtk_boxes.vtk_actor) # vtk_renderer.AddActor(vtk_point_cloud.vtk_actor) vtk_renderer.AddActor(vtk_voxel_grid.vtk_actor) vtk_renderer.AddActor(axes) vtk_renderer.SetBackground(0.2, 0.3, 0.4) # Setup Camera current_cam = vtk_renderer.GetActiveCamera() current_cam.Pitch(170.0) current_cam.Roll(180.0) # Zooms out to fit all points on screen vtk_renderer.ResetCamera() # Zoom in slightly current_cam.Zoom(2.5) # Reset the clipping range to show all points vtk_renderer.ResetCameraClippingRange() # Setup Render Window vtk_render_window = vtk.vtkRenderWindow() vtk_render_window.SetWindowName("Anchors") vtk_render_window.SetSize(900, 500) vtk_render_window.AddRenderer(vtk_renderer) # Setup custom interactor style, which handles mouse and key events vtk_render_window_interactor = vtk.vtkRenderWindowInteractor() vtk_render_window_interactor.SetRenderWindow(vtk_render_window) vtk_render_window_interactor.SetInteractorStyle( vtk.vtkInteractorStyleTrackballCamera()) # Render in VTK vtk_render_window.Render() vtk_render_window_interactor.Start() # Blocking
def main(): """This demo shows RPN proposals and MLOD predictions in 3D and 2D in image space. Given certain thresholds for proposals and predictions, it selects and draws the bounding boxes on the image sample. It goes through the entire proposal and prediction samples for the given dataset split. The proposals, overlaid, and prediction images can be toggled on or off separately in the options section. The prediction score and IoU with ground truth can be toggled on or off as well, shown as (score, IoU) above the detection. """ dataset_config = DatasetBuilder.copy_config(DatasetBuilder.KITTI_VAL) ############################## # Options ############################## dataset_config.data_split = 'val' fig_size = (10, 6.1) rpn_score_threshold = 0.1 mlod_score_threshold = 0.1 # Flag for projecting the 3D boxes to image space # in tensor format (for testing purposes) test_img_tensor_projection = False gt_classes = ['Pedestrian', 'Cyclist'] # gt_classes = ['Pedestrian', 'Cyclist'] # Overwrite this to select a specific checkpoint global_step = 44000 checkpoint_name = 'mlod_fpn_people' # Drawing Toggles draw_proposals_separate = False draw_overlaid = False draw_predictions_separate = True # Show orientation for both GT and proposals/predictions draw_orientations_on_prop = False draw_orientations_on_pred = False # Draw 2D bounding boxes draw_projected_2d_boxes = False # Save images for samples with no detections save_empty_images = True draw_score = True draw_iou = False iou_3d = False ############################## # End of Options ############################## # Get the dataset dataset = DatasetBuilder.build_kitti_dataset(dataset_config) # Setup Paths predictions_dir = mlod.root_dir() + \ '/data/outputs/' + checkpoint_name + '/predictions' proposals_and_scores_dir = predictions_dir + \ '/proposals_and_scores/' + dataset.data_split predictions_and_scores_dir = predictions_dir + \ '/final_predictions_and_scores/' + dataset.data_split # Output images directories output_dir_base = predictions_dir + '/images_2d' # Get checkpoint step steps = os.listdir(proposals_and_scores_dir) steps.sort(key=int) print('Available steps: {}'.format(steps)) # Use latest checkpoint if no index provided if global_step is None: global_step = steps[-1] if draw_proposals_separate: prop_out_dir = output_dir_base + '/proposals/{}/{}/{}'.format( dataset.data_split, global_step, rpn_score_threshold) if not os.path.exists(prop_out_dir): os.makedirs(prop_out_dir) print('Proposal images saved to:', prop_out_dir) if draw_overlaid: overlaid_out_dir = output_dir_base + '/overlaid/{}/{}/{}'.format( dataset.data_split, global_step, mlod_score_threshold) if not os.path.exists(overlaid_out_dir): os.makedirs(overlaid_out_dir) print('Overlaid images saved to:', overlaid_out_dir) if draw_predictions_separate: pred_out_dir = output_dir_base + '/predictions/{}/{}/{}'.format( dataset.data_split, global_step, mlod_score_threshold) if not os.path.exists(pred_out_dir): os.makedirs(pred_out_dir) print('Prediction images saved to:', pred_out_dir) # Rolling average array of times for time estimation avg_time_arr_length = 10 last_times = np.repeat(time.time(), avg_time_arr_length) + \ np.arange(avg_time_arr_length) for sample_idx in range(dataset.num_samples): # Estimate time remaining with 5 slowest times start_time = time.time() last_times = np.roll(last_times, -1) last_times[-1] = start_time avg_time = np.mean(np.sort(np.diff(last_times))[-5:]) samples_remaining = dataset.num_samples - sample_idx est_time_left = avg_time * samples_remaining # Print progress and time remaining estimate sys.stdout.write('\rSaving {} / {}, Avg Time: {:.3f}s, ' 'Time Remaining: {:.2f}s'. format( sample_idx + 1, dataset.num_samples, avg_time, est_time_left)) sys.stdout.flush() sample_name = dataset.sample_names[sample_idx] img_idx = int(sample_name) ############################## # Proposals ############################## if draw_proposals_separate or draw_overlaid: # Load proposals from files proposals_file_path = proposals_and_scores_dir + \ "/{}/{}.txt".format(global_step, sample_name) if not os.path.exists(proposals_file_path): print('Sample {}: No proposals, skipping'.format(sample_name)) continue print('Sample {}: Drawing proposals'.format(sample_name)) proposals_and_scores = np.loadtxt(proposals_file_path) proposal_boxes_3d = proposals_and_scores[:, 0:7] proposal_scores = proposals_and_scores[:, 7] # Apply score mask to proposals score_mask = proposal_scores > rpn_score_threshold proposal_boxes_3d = proposal_boxes_3d[score_mask] proposal_scores = proposal_scores[score_mask] proposal_objs = \ [box_3d_encoder.box_3d_to_object_label(proposal, obj_type='Proposal') for proposal in proposal_boxes_3d] ############################## # Predictions ############################## if draw_predictions_separate or draw_overlaid: predictions_file_path = predictions_and_scores_dir + \ "/{}/{}.txt".format(global_step, sample_name) if not os.path.exists(predictions_file_path): continue # Load predictions from files predictions_and_scores = np.loadtxt( predictions_and_scores_dir + "/{}/{}.txt".format(global_step, sample_name)) prediction_boxes_3d = predictions_and_scores[:, 0:7] prediction_scores = predictions_and_scores[:, 7] prediction_class_indices = predictions_and_scores[:, 8] # process predictions only if we have any predictions left after # masking if len(prediction_boxes_3d) > 0: # Apply score mask mlod_score_mask = prediction_scores >= 0.1 mlod_show_mask = mlod_score_mask prediction_boxes_3d = prediction_boxes_3d[mlod_show_mask] prediction_scores = prediction_scores[mlod_show_mask] prediction_class_indices = \ prediction_class_indices[mlod_show_mask] # # Swap l, w for predictions where w > l # swapped_indices = \ # prediction_boxes_3d[:, 4] > prediction_boxes_3d[:, 3] # prediction_boxes_3d = np.copy(prediction_boxes_3d) # prediction_boxes_3d[swapped_indices, 3] = \ # prediction_boxes_3d[swapped_indices, 4] # prediction_boxes_3d[swapped_indices, 4] = \ # prediction_boxes_3d[swapped_indices, 3] ############################## # Ground Truth ############################## # Get ground truth labels dataset.has_labels = False if dataset.has_labels: gt_objects = obj_utils.read_labels(dataset.label_dir, img_idx) else: gt_objects = [] # Filter objects to desired difficulty filtered_gt_objs = dataset.kitti_utils.filter_labels( gt_objects, classes=gt_classes) boxes2d, _, _ = obj_utils.build_bbs_from_objects( filtered_gt_objs, class_needed=gt_classes) image_path = dataset.get_rgb_image_path(sample_name) image = Image.open(image_path) image_size = image.size # Read the stereo calibration matrix for visualization stereo_calib = calib_utils.read_calibration(dataset.calib_dir, img_idx) calib_p2 = stereo_calib.p2 ############################## # Reformat and prepare to draw ############################## if draw_proposals_separate or draw_overlaid: proposals_as_anchors = box_3d_encoder.box_3d_to_anchor( proposal_boxes_3d) if test_img_tensor_projection: proposal_boxes = demo_utils.tf_project_to_image_space( proposals_as_anchors, calib_p2, image_size, img_idx) else: proposal_boxes, _ = anchor_projector.project_to_image_space( proposals_as_anchors, calib_p2, image_size) num_of_proposals = proposal_boxes_3d.shape[0] prop_fig, prop_2d_axes, prop_3d_axes = \ vis_utils.visualization(dataset.rgb_image_dir, img_idx, display=False) draw_proposals(filtered_gt_objs, calib_p2, num_of_proposals, proposal_objs, proposal_boxes, prop_2d_axes, prop_3d_axes, draw_orientations_on_prop) if draw_proposals_separate: # Save just the proposals filename = prop_out_dir + '/' + sample_name + '.png' plt.savefig(filename) if not draw_overlaid: plt.close(prop_fig) if draw_overlaid or draw_predictions_separate: if len(prediction_boxes_3d) > 0: # Project the 3D box predictions to image space image_filter = [] final_boxes_2d = [] for i in range(len(prediction_boxes_3d)): box_3d = prediction_boxes_3d[i, 0:7] img_box = box_3d_projector.project_to_image_space( box_3d, calib_p2, truncate=True, image_size=image_size, discard_before_truncation=False) if img_box is not None: image_filter.append(True) final_boxes_2d.append(img_box) else: image_filter.append(False) final_boxes_2d = np.asarray(final_boxes_2d) final_prediction_boxes_3d = prediction_boxes_3d[image_filter] final_scores = prediction_scores[image_filter] final_class_indices = prediction_class_indices[image_filter] num_of_predictions = final_boxes_2d.shape[0] # Convert to objs final_prediction_objs = \ [box_3d_encoder.box_3d_to_object_label( prediction, obj_type='Prediction') for prediction in final_prediction_boxes_3d] for (obj, score) in zip(final_prediction_objs, final_scores): obj.score = score else: if save_empty_images: pred_fig, pred_2d_axes, pred_3d_axes = \ vis_utils.visualization(dataset.rgb_image_dir, img_idx, display=False, fig_size=fig_size) filename = pred_out_dir + '/' + sample_name + '.png' plt.savefig(filename) plt.close(pred_fig) continue if draw_overlaid: # Overlay prediction boxes on image draw_predictions(filtered_gt_objs, calib_p2, num_of_predictions, final_prediction_objs, final_class_indices, final_boxes_2d, prop_2d_axes, prop_3d_axes, draw_score, draw_iou, gt_classes, draw_orientations_on_pred, iou_3d) filename = overlaid_out_dir + '/' + sample_name + '.png' plt.savefig(filename) plt.close(prop_fig) if draw_predictions_separate: # Now only draw prediction boxes on images # on a new figure handler if draw_projected_2d_boxes: pred_fig, pred_2d_axes, pred_3d_axes = \ vis_utils.visualization(dataset.rgb_image_dir, img_idx, display=False, fig_size=fig_size) draw_predictions(filtered_gt_objs, calib_p2, num_of_predictions, final_prediction_objs, final_class_indices, final_boxes_2d, pred_2d_axes, pred_3d_axes, draw_score, draw_iou, gt_classes, draw_orientations_on_pred, iou_3d) else: pred_fig, pred_3d_axes = \ vis_utils.visualize_single_plot( dataset.rgb_image_dir, img_idx, display=False) draw_3d_predictions(filtered_gt_objs, calib_p2, num_of_predictions, final_prediction_objs, final_class_indices, final_boxes_2d, pred_3d_axes, draw_score, draw_iou, gt_classes, draw_orientations_on_pred) filename = pred_out_dir + '/' + sample_name + '.png' plt.savefig(filename) plt.close(pred_fig) print('\nDone')
def np_box_3d_to_box_4c(box_3d, ground_plane): """Converts a single box_3d to box_4c Args: box_3d: box_3d (6,) ground_plane: ground plane coefficients (4,) Returns: box_4c (10,) """ format_checker.check_box_3d_format(box_3d) anchor = box_3d_encoder.box_3d_to_anchor(box_3d, ortho_rotate=True)[0] centroid_x = anchor[0] centroid_y = anchor[1] centroid_z = anchor[2] dim_x = anchor[3] dim_y = anchor[4] dim_z = anchor[5] # Create temporary box at (0, 0) for rotation half_dim_x = dim_x / 2 half_dim_z = dim_z / 2 # Box corners x_corners = np.asarray([half_dim_x, half_dim_x, -half_dim_x, -half_dim_x]) z_corners = np.array([half_dim_z, -half_dim_z, -half_dim_z, half_dim_z]) ry = box_3d[6] # Find nearest 90 degree half_pi = np.pi / 2 ortho_ry = np.round(ry / half_pi) * half_pi # Find rotation to make the box ortho aligned ry_diff = ry - ortho_ry # Create transformation matrix, including rotation and translation tr_mat = np.array([[np.cos(ry_diff), np.sin(ry_diff), centroid_x], [-np.sin(ry_diff), np.cos(ry_diff), centroid_z], [0, 0, 1]]) # Create a ones row ones_row = np.ones(x_corners.shape) # Append the column of ones to be able to multiply points_stacked = np.vstack([x_corners, z_corners, ones_row]) corners = np.matmul(tr_mat, points_stacked) # Discard the last row (ones) corners = corners[0:2] # Calculate height off ground plane ground_y = geometry_utils.calculate_plane_point( ground_plane, [centroid_x, None, centroid_z])[1] h1 = ground_y - centroid_y h2 = h1 + dim_y # Stack into (10,) ndarray box_4c = np.hstack([corners.flatten(), h1, h2]) return box_4c
def main(): """ This demo shows example mini batch info for full MlodModel training. This includes ground truth, ortho rotated ground truth, negative proposal anchors, positive proposal anchors, and a sampled mini batch. The 2D iou can be modified to show the effect of changing the iou threshold for mini batch sampling. In order to let this demo run without training an RPN, the proposals shown are being read from a text file. Keys: F1: Toggle ground truth F2: Toggle ortho rotated ground truth F3: Toggle negative proposal anchors F4: Toggle positive proposal anchors F5: Toggle mini batch anchors """ ############################## # Options ############################## # Config file folder, default (<mlod_root>/data/outputs/<checkpoint_name>) config_dir = None # checkpoint_name = None checkpoint_name = 'mlod_exp_example' data_split = 'val_half' # global_step = None global_step = 100000 # # # Cars # # # # sample_name = "000050" sample_name = "000104" # sample_name = "000764" # # # People # # # # val_half # sample_name = '000001' # Hard, 1 far cyc # sample_name = '000005' # Easy, 1 ped # sample_name = '000122' # Easy, 1 cyc # sample_name = '000134' # Hard, lots of people # sample_name = '000167' # Medium, 1 ped, 2 cycs # sample_name = '000187' # Medium, 1 ped on left # sample_name = '000381' # Easy, 1 ped # sample_name = '000398' # Easy, 1 ped # sample_name = '000401' # Hard, obscured peds # sample_name = '000407' # Easy, 1 ped # sample_name = '000448' # Hard, several far people # sample_name = '000486' # Hard 2 obscured peds # sample_name = '000509' # Easy, 1 ped # sample_name = '000718' # Hard, lots of people # sample_name = '002216' # Easy, 1 cyc mini_batch_size = 512 neg_proposal_2d_iou_hi = 0.6 pos_proposal_2d_iou_lo = 0.65 bkg_proposals_line_width = 0.5 neg_proposals_line_width = 0.5 mid_proposals_line_width = 0.5 pos_proposals_line_width = 1.0 ############################## # End of Options ############################## img_idx = int(sample_name) print("Showing mini batch for sample {}".format(sample_name)) # Read proposals from file if checkpoint_name is None: # Use VAL Dataset dataset = DatasetBuilder.build_kitti_dataset(DatasetBuilder.KITTI_VAL) # Load demo proposals proposals_and_scores_dir = mlod.top_dir() + \ '/demos/data/predictions/' + checkpoint_name + \ '/proposals_and_scores/' + dataset.data_split else: if config_dir is None: config_dir = mlod.root_dir() + '/data/outputs/' + checkpoint_name # Parse experiment config pipeline_config_file = \ config_dir + '/' + checkpoint_name + '.config' _, _, _, dataset_config = \ config_builder_util.get_configs_from_pipeline_file( pipeline_config_file, is_training=False) dataset_config.data_split = data_split dataset = DatasetBuilder.build_kitti_dataset(dataset_config, use_defaults=False) # Overwrite mini_batch_utils = dataset.kitti_utils.mini_batch_utils mini_batch_utils.mlod_neg_iou_range[1] = neg_proposal_2d_iou_hi mini_batch_utils.mlod_pos_iou_range[0] = pos_proposal_2d_iou_lo # Load proposals from outputs folder proposals_and_scores_dir = mlod.root_dir() + \ '/data/outputs/' + checkpoint_name + \ '/predictions/proposals_and_scores/' + dataset.data_split # Get checkpoint step steps = os.listdir(proposals_and_scores_dir) steps.sort(key=int) print('Available steps: {}'.format(steps)) # Use latest checkpoint if no index provided if global_step is None: global_step = steps[-1] proposals_and_scores = np.loadtxt( proposals_and_scores_dir + "/{}/{}.txt".format(global_step, sample_name)) proposal_boxes_3d = proposals_and_scores[:, 0:7] proposal_anchors = box_3d_encoder.box_3d_to_anchor(proposal_boxes_3d) # Get filtered ground truth obj_labels = obj_utils.read_labels(dataset.label_dir, img_idx) filtered_objs = dataset.kitti_utils.filter_labels(obj_labels) # Convert ground truth to anchors gt_boxes_3d = np.asarray([ box_3d_encoder.object_label_to_box_3d(obj_label) for obj_label in filtered_objs ]) gt_anchors = box_3d_encoder.box_3d_to_anchor(gt_boxes_3d, ortho_rotate=True) # Ortho rotate ground truth gt_ortho_boxes_3d = box_3d_encoder.anchors_to_box_3d(gt_anchors) gt_ortho_objs = [ box_3d_encoder.box_3d_to_object_label(box_3d, obj_type='OrthoGt') for box_3d in gt_ortho_boxes_3d ] # Project gt and anchors into BEV gt_bev_anchors, _ = \ anchor_projector.project_to_bev(gt_anchors, dataset.kitti_utils.bev_extents) bev_anchors, _ = \ anchor_projector.project_to_bev(proposal_anchors, dataset.kitti_utils.bev_extents) # Reorder boxes into (y1, x1, y2, x2) order gt_bev_anchors_tf_order = anchor_projector.reorder_projected_boxes( gt_bev_anchors) bev_anchors_tf_order = anchor_projector.reorder_projected_boxes( bev_anchors) # Convert to box_list format for iou calculation gt_anchor_box_list = box_list.BoxList( tf.cast(gt_bev_anchors_tf_order, tf.float32)) anchor_box_list = box_list.BoxList( tf.cast(bev_anchors_tf_order, tf.float32)) # Get IoU for every anchor tf_all_ious = box_list_ops.iou(gt_anchor_box_list, anchor_box_list) valid_ious = True # Make sure the calculated IoUs contain values. Since its a [N, M] # tensor, if there are no gt's for instance, that entry will be zero. if tf_all_ious.shape[0] == 0 or tf_all_ious.shape[1] == 0: print('#################################################') print('Warning: This sample does not contain valid IoUs') print('#################################################') valid_ious = False if valid_ious: tf_max_ious = tf.reduce_max(tf_all_ious, axis=0) tf_max_iou_indices = tf.argmax(tf_all_ious, axis=0) # Sample an RPN mini batch from the non empty anchors mini_batch_utils = dataset.kitti_utils.mini_batch_utils # Overwrite mini batch size and sample a mini batch mini_batch_utils.mlod_mini_batch_size = mini_batch_size mb_mask_tf, _ = mini_batch_utils.sample_mlod_mini_batch(tf_max_ious) # Create a session config = tf.ConfigProto() config.gpu_options.allow_growth = True sess = tf.Session(config=config) # Run the graph to calculate ious for every proposal and # to get the mini batch mask all_ious, max_ious, max_iou_indices = sess.run( [tf_all_ious, tf_max_ious, tf_max_iou_indices]) mb_mask = sess.run(mb_mask_tf) mb_anchors = proposal_anchors[mb_mask] mb_anchor_boxes_3d = box_3d_encoder.anchors_to_box_3d(mb_anchors) mb_anchor_ious = max_ious[mb_mask] else: # We have no valid IoU's, so assume all IoUs are zeros # and the mini-batch contains all the anchors since we cannot # mask without IoUs. max_ious = np.zeros(proposal_boxes_3d.shape[0]) mb_anchor_ious = max_ious mb_anchors = proposal_anchors mb_anchor_boxes_3d = box_3d_encoder.anchors_to_box_3d(mb_anchors) # Create list of positive/negative proposals based on iou pos_proposal_objs = [] mid_proposal_objs = [] neg_proposal_objs = [] bkg_proposal_objs = [] for i in range(len(proposal_boxes_3d)): box_3d = proposal_boxes_3d[i] if max_ious[i] == 0.0: # Background proposals bkg_proposal_objs.append( box_3d_encoder.box_3d_to_object_label( box_3d, obj_type='BackgroundProposal')) elif max_ious[i] < neg_proposal_2d_iou_hi: # Negative proposals neg_proposal_objs.append( box_3d_encoder.box_3d_to_object_label( box_3d, obj_type='NegativeProposal')) elif max_ious[i] < pos_proposal_2d_iou_lo: # Middle proposals (in between negative and positive) mid_proposal_objs.append( box_3d_encoder.box_3d_to_object_label( box_3d, obj_type='MiddleProposal')) elif max_ious[i] <= 1.0: # Positive proposals pos_proposal_objs.append( box_3d_encoder.box_3d_to_object_label( box_3d, obj_type='PositiveProposal')) else: raise ValueError('Invalid IoU > 1.0') print('{} bkg, {} neg, {} mid, {} pos proposals:'.format( len(bkg_proposal_objs), len(neg_proposal_objs), len(mid_proposal_objs), len(pos_proposal_objs))) # Convert the mini_batch anchors to object list mb_obj_list = [] for i in range(len(mb_anchor_ious)): if valid_ious and (mb_anchor_ious[i] > mini_batch_utils.mlod_pos_iou_range[0]): obj_type = "Positive" else: obj_type = "Negative" obj = box_3d_encoder.box_3d_to_object_label(mb_anchor_boxes_3d[i], obj_type) mb_obj_list.append(obj) # Point cloud image = cv2.imread(dataset.get_rgb_image_path(sample_name)) points, point_colours = demo_utils.get_filtered_pc_and_colours( dataset, image, img_idx) # Visualize from here vis_utils.visualization(dataset.rgb_image_dir, img_idx) plt.show(block=False) # VtkPointCloud vtk_point_cloud = VtkPointCloud() vtk_point_cloud.set_points(points, point_colours) # VtkAxes axes = vtk.vtkAxesActor() axes.SetTotalLength(5, 5, 5) # VtkBoxes for ground truth vtk_gt_boxes = VtkBoxes() vtk_gt_boxes.set_objects(filtered_objs, COLOUR_SCHEME) # VtkBoxes for ortho ground truth vtk_gt_ortho_boxes = VtkBoxes() vtk_gt_ortho_boxes.set_objects(gt_ortho_objs, COLOUR_SCHEME) # VtkBoxes for background proposals vtk_bkg_proposal_boxes = VtkBoxes() vtk_bkg_proposal_boxes.set_objects(bkg_proposal_objs, COLOUR_SCHEME) vtk_bkg_proposal_boxes.set_line_width(bkg_proposals_line_width) # VtkBoxes for negative proposals vtk_neg_proposal_boxes = VtkBoxes() vtk_neg_proposal_boxes.set_objects(neg_proposal_objs, COLOUR_SCHEME) vtk_neg_proposal_boxes.set_line_width(neg_proposals_line_width) # VtkBoxes for middle proposals vtk_mid_proposal_boxes = VtkBoxes() vtk_mid_proposal_boxes.set_objects(mid_proposal_objs, COLOUR_SCHEME) vtk_mid_proposal_boxes.set_line_width(mid_proposals_line_width) # VtkBoxes for positive proposals vtk_pos_proposal_boxes = VtkBoxes() vtk_pos_proposal_boxes.set_objects(pos_proposal_objs, COLOUR_SCHEME) vtk_pos_proposal_boxes.set_line_width(pos_proposals_line_width) # Create VtkBoxes for mini batch anchors vtk_mb_boxes = VtkBoxes() vtk_mb_boxes.set_objects(mb_obj_list, COLOUR_SCHEME) # Create Voxel Grid Renderer in bottom half vtk_renderer = vtk.vtkRenderer() vtk_renderer.SetBackground(0.2, 0.3, 0.4) # Add actors vtk_renderer.AddActor(axes) vtk_renderer.AddActor(vtk_point_cloud.vtk_actor) vtk_renderer.AddActor(vtk_gt_boxes.vtk_actor) vtk_renderer.AddActor(vtk_gt_ortho_boxes.vtk_actor) vtk_renderer.AddActor(vtk_bkg_proposal_boxes.vtk_actor) vtk_renderer.AddActor(vtk_neg_proposal_boxes.vtk_actor) vtk_renderer.AddActor(vtk_mid_proposal_boxes.vtk_actor) vtk_renderer.AddActor(vtk_pos_proposal_boxes.vtk_actor) vtk_renderer.AddActor(vtk_mb_boxes.vtk_actor) # Setup Camera current_cam = vtk_renderer.GetActiveCamera() current_cam.Pitch(160.0) current_cam.Roll(180.0) # Zooms out to fit all points on screen vtk_renderer.ResetCamera() # Zoom in slightly current_cam.Zoom(2.5) # Reset the clipping range to show all points vtk_renderer.ResetCameraClippingRange() # Setup Render Window vtk_render_window = vtk.vtkRenderWindow() vtk_render_window.SetWindowName("MLOD Mini Batch") vtk_render_window.SetSize(900, 500) vtk_render_window.AddRenderer(vtk_renderer) # Setup custom interactor style, which handles mouse and key events vtk_render_window_interactor = vtk.vtkRenderWindowInteractor() vtk_render_window_interactor.SetRenderWindow(vtk_render_window) vtk_render_window_interactor.SetInteractorStyle( vis_utils.ToggleActorsInteractorStyle([ vtk_gt_boxes.vtk_actor, vtk_gt_ortho_boxes.vtk_actor, vtk_bkg_proposal_boxes.vtk_actor, vtk_neg_proposal_boxes.vtk_actor, vtk_mid_proposal_boxes.vtk_actor, vtk_pos_proposal_boxes.vtk_actor, vtk_mb_boxes.vtk_actor, ])) # Render in VTK vtk_render_window.Render() vtk_render_window_interactor.Start()
def main(): """ Visualization of 3D grid anchor generation, showing 2D projections in BEV and image space, and a 3D display of the anchors """ dataset_config = DatasetBuilder.copy_config(DatasetBuilder.KITTI_TRAIN) dataset_config.num_clusters[0] = 1 dataset = DatasetBuilder.build_kitti_dataset(dataset_config) label_cluster_utils = LabelClusterUtils(dataset) clusters, _ = label_cluster_utils.get_clusters() # Options img_idx = 1 # fake_clusters = np.array([[5, 4, 3], [6, 5, 4]]) # fake_clusters = np.array([[3, 3, 3], [4, 4, 4]]) fake_clusters = np.array([[4, 2, 3]]) fake_anchor_stride = [5.0, 5.0] ground_plane = [0, -1, 0, 1.72] anchor_3d_generator = grid_anchor_3d_generator.GridAnchor3dGenerator() area_extents = np.array([[-40, 40], [-5, 5], [0, 70]]) # Generate anchors for cars only start_time = time.time() anchor_boxes_3d = anchor_3d_generator.generate( area_3d=dataset.kitti_utils.area_extents, anchor_3d_sizes=fake_clusters, anchor_stride=fake_anchor_stride, ground_plane=ground_plane) all_anchors = box_3d_encoder.box_3d_to_anchor(anchor_boxes_3d) end_time = time.time() print("Anchors generated in {} s".format(end_time - start_time)) # Project into bev bev_boxes, bev_normalized_boxes = \ anchor_projector.project_to_bev(all_anchors, area_extents[[0, 2]]) bev_fig, (bev_axes, bev_normalized_axes) = \ plt.subplots(1, 2, figsize=(16, 7)) bev_axes.set_xlim(0, 80) bev_axes.set_ylim(70, 0) bev_normalized_axes.set_xlim(0, 1.0) bev_normalized_axes.set_ylim(1, 0.0) plt.show(block=False) for box in bev_boxes: box_w = box[2] - box[0] box_h = box[3] - box[1] rect = patches.Rectangle((box[0], box[1]), box_w, box_h, linewidth=2, edgecolor='b', facecolor='none') bev_axes.add_patch(rect) for normalized_box in bev_normalized_boxes: box_w = normalized_box[2] - normalized_box[0] box_h = normalized_box[3] - normalized_box[1] rect = patches.Rectangle((normalized_box[0], normalized_box[1]), box_w, box_h, linewidth=2, edgecolor='b', facecolor='none') bev_normalized_axes.add_patch(rect) rgb_fig, rgb_2d_axes, rgb_3d_axes = \ vis_utils.visualization(dataset.rgb_image_dir, img_idx) plt.show(block=False) image_path = dataset.get_rgb_image_path(dataset.sample_names[img_idx]) image_shape = np.array(Image.open(image_path)).shape stereo_calib_p2 = calib_utils.read_calibration(dataset.calib_dir, img_idx).p2 start_time = time.time() rgb_boxes, rgb_normalized_boxes = \ anchor_projector.project_to_image_space(all_anchors, stereo_calib_p2, image_shape) end_time = time.time() print("Anchors projected in {} s".format(end_time - start_time)) # Read the stereo calibration matrix for visualization stereo_calib = calib_utils.read_calibration(dataset.calib_dir, 0) p = stereo_calib.p2 # Overlay boxes on images anchor_objects = [] for anchor_idx in range(len(anchor_boxes_3d)): anchor_box_3d = anchor_boxes_3d[anchor_idx] obj_label = box_3d_encoder.box_3d_to_object_label(anchor_box_3d) # Append to a list for visualization in VTK later anchor_objects.append(obj_label) # Draw 3D boxes vis_utils.draw_box_3d(rgb_3d_axes, obj_label, p) # Draw 2D boxes rgb_box_2d = rgb_boxes[anchor_idx] box_x1 = rgb_box_2d[0] box_y1 = rgb_box_2d[1] box_w = rgb_box_2d[2] - box_x1 box_h = rgb_box_2d[3] - box_y1 rect = patches.Rectangle((box_x1, box_y1), box_w, box_h, linewidth=2, edgecolor='b', facecolor='none') rgb_2d_axes.add_patch(rect) if anchor_idx % 32 == 0: rgb_fig.canvas.draw() plt.show(block=False) # Create VtkGroundPlane for ground plane visualization vtk_ground_plane = VtkGroundPlane() vtk_ground_plane.set_plane(ground_plane, area_extents[[0, 2]]) # Create VtkAxes axes = vtk.vtkAxesActor() axes.SetTotalLength(5, 5, 5) # Create VtkBoxes for boxes vtk_boxes = VtkBoxes() vtk_boxes.set_objects(anchor_objects, vtk_boxes.COLOUR_SCHEME_KITTI) # Create Voxel Grid Renderer in bottom half vtk_renderer = vtk.vtkRenderer() vtk_renderer.AddActor(vtk_boxes.vtk_actor) vtk_renderer.AddActor(vtk_ground_plane.vtk_actor) vtk_renderer.AddActor(axes) vtk_renderer.SetBackground(0.2, 0.3, 0.4) # Setup Camera current_cam = vtk_renderer.GetActiveCamera() current_cam.Pitch(170.0) current_cam.Roll(180.0) # Zooms out to fit all points on screen vtk_renderer.ResetCamera() # Zoom in slightly current_cam.Zoom(2.5) # Reset the clipping range to show all points vtk_renderer.ResetCameraClippingRange() # Setup Render Window vtk_render_window = vtk.vtkRenderWindow() vtk_render_window.SetWindowName("Anchors") vtk_render_window.SetSize(900, 500) vtk_render_window.AddRenderer(vtk_renderer) # Setup custom interactor style, which handles mouse and key events vtk_render_window_interactor = vtk.vtkRenderWindowInteractor() vtk_render_window_interactor.SetRenderWindow(vtk_render_window) vtk_render_window_interactor.SetInteractorStyle( vtk.vtkInteractorStyleTrackballCamera()) # Render in VTK vtk_render_window.Render() vtk_render_window_interactor.Start() # Blocking
def main(): """Flip RPN Mini Batch Visualization of the mini batch anchors for RpnModel training. Keys: F1: Toggle mini batch anchors F2: Flipped """ anchor_colour_scheme = { "Car": (255, 0, 0), # Red "Pedestrian": (255, 150, 50), # Orange "Cyclist": (150, 50, 100), # Purple "DontCare": (255, 255, 255), # White "Anchor": (150, 150, 150), # Gray "Regressed Anchor": (255, 255, 0), # Yellow "Positive": (0, 255, 255), # Teal "Negative": (255, 0, 255) # Purple } dataset_config_path = mlod.root_dir() + \ '/configs/mb_rpn_demo_cars.config' # dataset_config_path = mlod.root_dir() + \ # '/configs/mb_rpn_demo_people.config' ############################## # Options ############################## # # # Random sample # # # sample_name = None # # # Cars # # # # sample_name = "000001" # sample_name = "000050" # sample_name = "000104" # sample_name = "000112" # sample_name = "000169" # sample_name = "000191" sample_name = "003801" # # # Pedestrians # # # # sample_name = "000000" # sample_name = "000011" # sample_name = "000015" # sample_name = "000028" # sample_name = "000035" # sample_name = "000134" # sample_name = "000167" # sample_name = '000379' # sample_name = '000381' # sample_name = '000397' # sample_name = '000398' # sample_name = '000401' # sample_name = '000407' # sample_name = '000486' # sample_name = '000509' # # Cyclists # # # # sample_name = '000122' # sample_name = '000448' # # # Multiple classes # # # # sample_name = "000764" ############################## # End of Options ############################## # Create Dataset dataset = DatasetBuilder.load_dataset_from_config(dataset_config_path) # Random sample if sample_name is None: sample_idx = np.random.randint(0, dataset.num_samples) sample_name = dataset.sample_list[sample_idx] anchor_strides = dataset.kitti_utils.anchor_strides img_idx = int(sample_name) print("Showing mini batch for sample {}".format(sample_name)) image = cv2.imread(dataset.get_rgb_image_path(sample_name)) image_shape = [image.shape[1], image.shape[0]] # KittiUtils class dataset_utils = dataset.kitti_utils ground_plane = obj_utils.get_road_plane(img_idx, dataset.planes_dir) point_cloud = obj_utils.get_depth_map_point_cloud(img_idx, dataset.calib_dir, dataset.depth_dir, image_shape) points = point_cloud.T # Grab ground truth ground_truth_list = obj_utils.read_labels(dataset.label_dir, img_idx) ground_truth_list = dataset_utils.filter_labels(ground_truth_list) stereo_calib_p2 = calib_utils.read_calibration(dataset.calib_dir, img_idx).p2 ############################## # Flip sample info ############################## start_time = time.time() flipped_image = kitti_aug.flip_image(image) flipped_point_cloud = kitti_aug.flip_point_cloud(point_cloud) flipped_gt_list = [ kitti_aug.flip_label_in_3d_only(obj) for obj in ground_truth_list ] flipped_ground_plane = kitti_aug.flip_ground_plane(ground_plane) flipped_calib_p2 = kitti_aug.flip_stereo_calib_p2(stereo_calib_p2, image_shape) print('flip sample', time.time() - start_time) flipped_points = flipped_point_cloud.T point_colours = vis_utils.project_img_to_point_cloud( points, image, dataset.calib_dir, img_idx) ############################## # Generate anchors ############################## clusters, _ = dataset.get_cluster_info() anchor_generator = grid_anchor_3d_generator.GridAnchor3dGenerator() # Read mini batch info anchors_info = dataset_utils.get_anchors_info(sample_name) all_anchor_boxes_3d = [] all_ious = [] all_offsets = [] for class_idx in range(len(dataset.classes)): anchor_boxes_3d = anchor_generator.generate( area_3d=dataset.kitti_utils.area_extents, anchor_3d_sizes=clusters[class_idx], anchor_stride=anchor_strides[class_idx], ground_plane=ground_plane) if len(anchors_info[class_idx]) > 0: indices, ious, offsets, classes = anchors_info[class_idx] # Get non empty anchors from the indices non_empty_anchor_boxes_3d = anchor_boxes_3d[indices] all_anchor_boxes_3d.extend(non_empty_anchor_boxes_3d) all_ious.extend(ious) all_offsets.extend(offsets) if not len(all_anchor_boxes_3d) > 0: # Exit early if anchors_info is empty print("No anchors, Please try a different sample") return # Convert to ndarrays all_anchor_boxes_3d = np.asarray(all_anchor_boxes_3d) all_ious = np.asarray(all_ious) all_offsets = np.asarray(all_offsets) ############################## # Flip anchors ############################## start_time = time.time() # Flip anchors and offsets flipped_anchor_boxes_3d = kitti_aug.flip_boxes_3d(all_anchor_boxes_3d, flip_ry=False) all_offsets[:, 0] = -all_offsets[:, 0] print('flip anchors and offsets', time.time() - start_time) # Overwrite with flipped things all_anchor_boxes_3d = flipped_anchor_boxes_3d points = flipped_points ground_truth_list = flipped_gt_list ground_plane = flipped_ground_plane ############################## # Mini batch sampling ############################## # Sample an RPN mini batch from the non empty anchors mini_batch_utils = dataset.kitti_utils.mini_batch_utils mb_mask_tf, _ = mini_batch_utils.sample_rpn_mini_batch(all_ious) config = tf.ConfigProto() config.gpu_options.allow_growth = True sess = tf.Session(config=config) mb_mask = sess.run(mb_mask_tf) mb_anchor_boxes_3d = all_anchor_boxes_3d[mb_mask] mb_anchor_ious = all_ious[mb_mask] mb_anchor_offsets = all_offsets[mb_mask] # ObjectLabel list that hold all boxes to visualize obj_list = [] # Convert the mini_batch anchors to object list for i in range(len(mb_anchor_boxes_3d)): if mb_anchor_ious[i] > mini_batch_utils.rpn_pos_iou_range[0]: obj_type = "Positive" else: obj_type = "Negative" obj = box_3d_encoder.box_3d_to_object_label(mb_anchor_boxes_3d[i], obj_type) obj_list.append(obj) # Convert all non-empty anchors to object list non_empty_anchor_objs = \ [box_3d_encoder.box_3d_to_object_label( anchor_box_3d, obj_type='Anchor') for anchor_box_3d in all_anchor_boxes_3d] ############################## # Regress Positive Anchors ############################## # Convert anchor_boxes_3d to anchors and apply offsets mb_pos_mask = mb_anchor_ious > mini_batch_utils.rpn_pos_iou_range[0] mb_pos_anchor_boxes_3d = mb_anchor_boxes_3d[mb_pos_mask] mb_pos_anchor_offsets = mb_anchor_offsets[mb_pos_mask] mb_pos_anchors = box_3d_encoder.box_3d_to_anchor(mb_pos_anchor_boxes_3d) regressed_pos_anchors = anchor_encoder.offset_to_anchor( mb_pos_anchors, mb_pos_anchor_offsets) # Convert regressed anchors to ObjectLabels for visualization regressed_anchor_boxes_3d = box_3d_encoder.anchors_to_box_3d( regressed_pos_anchors, fix_lw=True) regressed_anchor_objs = \ [box_3d_encoder.box_3d_to_object_label( box_3d, obj_type='Regressed Anchor') for box_3d in regressed_anchor_boxes_3d] ############################## # Visualization ############################## cv2.imshow('{} flipped'.format(sample_name), flipped_image) cv2.waitKey() # Create VtkAxes axes = vtk.vtkAxesActor() axes.SetTotalLength(5, 5, 5) # Create VtkBoxes for mini batch anchors vtk_pos_anchor_boxes = VtkBoxes() vtk_pos_anchor_boxes.set_objects(obj_list, anchor_colour_scheme) # VtkBoxes for non empty anchors vtk_non_empty_anchors = VtkBoxes() vtk_non_empty_anchors.set_objects(non_empty_anchor_objs, anchor_colour_scheme) vtk_non_empty_anchors.set_line_width(0.1) # VtkBoxes for regressed anchors vtk_regressed_anchors = VtkBoxes() vtk_regressed_anchors.set_objects(regressed_anchor_objs, anchor_colour_scheme) vtk_regressed_anchors.set_line_width(5.0) # Create VtkBoxes for ground truth vtk_gt_boxes = VtkBoxes() vtk_gt_boxes.set_objects(ground_truth_list, anchor_colour_scheme, show_orientations=True) vtk_point_cloud = VtkPointCloud() vtk_point_cloud.set_points(points, point_colours) vtk_ground_plane = VtkGroundPlane() vtk_ground_plane.set_plane(ground_plane, dataset.kitti_utils.bev_extents) # Create Voxel Grid Renderer in bottom half vtk_renderer = vtk.vtkRenderer() vtk_renderer.AddActor(vtk_point_cloud.vtk_actor) vtk_renderer.AddActor(vtk_non_empty_anchors.vtk_actor) vtk_renderer.AddActor(vtk_pos_anchor_boxes.vtk_actor) vtk_renderer.AddActor(vtk_regressed_anchors.vtk_actor) vtk_renderer.AddActor(vtk_gt_boxes.vtk_actor) vtk_renderer.AddActor(vtk_ground_plane.vtk_actor) vtk_renderer.AddActor(axes) vtk_renderer.SetBackground(0.2, 0.3, 0.4) # Setup Camera current_cam = vtk_renderer.GetActiveCamera() current_cam.Pitch(160.0) current_cam.Roll(180.0) # Zooms out to fit all points on screen vtk_renderer.ResetCamera() # Zoom in slightly current_cam.Zoom(2.5) # Reset the clipping range to show all points vtk_renderer.ResetCameraClippingRange() # Setup Render Window vtk_render_window = vtk.vtkRenderWindow() vtk_render_window.SetWindowName("RPN Mini Batch") vtk_render_window.SetSize(900, 500) vtk_render_window.AddRenderer(vtk_renderer) # Setup custom interactor style, which handles mouse and key events vtk_render_window_interactor = vtk.vtkRenderWindowInteractor() vtk_render_window_interactor.SetRenderWindow(vtk_render_window) vtk_render_window_interactor.SetInteractorStyle( vis_utils.ToggleActorsInteractorStyle([ vtk_non_empty_anchors.vtk_actor, vtk_pos_anchor_boxes.vtk_actor, vtk_regressed_anchors.vtk_actor, vtk_ground_plane.vtk_actor, ])) # Render in VTK vtk_render_window.Render() vtk_render_window_interactor.Start()
def main(): """ Simple demo script for debugging integral images with visualization """ anchor_colour_scheme = {"Anchor": (0, 0, 255)} # Blue dataset = DatasetBuilder.build_kitti_dataset(DatasetBuilder.KITTI_TRAIN) label_cluster_utils = LabelClusterUtils(dataset) clusters, _ = label_cluster_utils.get_clusters() area_extents = np.array([[0, 2], [-1, 0.], [0, 2]]) boxes_3d = np.array([ [2, 0, 1, 1, 1, 1, 0], [1, 0, 2, 1, 1, 1, 0], ]) xyz = np.array([[0.5, -0.01, 1.1], [1.5, -0.01, 1.1], [0.5, -0.01, 1.6], [1.5, -0.01, 1.6], [0.5, -0.49, 1.1], [1.5, -0.49, 1.1], [0.5, -0.51, 1.6], [1.5, -0.51, 1.6] ]) vx_grid_3d = voxel_grid.VoxelGrid() vx_grid_3d.voxelize(xyz, 0.1, area_extents) anchors = box_3d_encoder.box_3d_to_anchor(boxes_3d) # Filter the boxes here! start_time = time.time() empty_filter = anchor_filter.get_empty_anchor_filter(anchors=anchors, voxel_grid_3d=vx_grid_3d, density_threshold=1) boxes_3d = boxes_3d[empty_filter] end_time = time.time() print("Anchors filtered in {} s".format(end_time - start_time)) box_objects = [] for box_idx in range(len(boxes_3d)): box = boxes_3d[box_idx] obj_label = box_3d_encoder.box_3d_to_object_label(box, 'Anchor') # Append to a list for visualization in VTK later box_objects.append(obj_label) # Create VtkAxes axes = vtk.vtkAxesActor() axes.SetTotalLength(5, 5, 5) # Create VtkBoxes for boxes vtk_boxes = VtkBoxes() vtk_boxes.set_objects(box_objects, anchor_colour_scheme) vtk_point_cloud = VtkPointCloud() vtk_point_cloud.set_points(xyz) vtk_voxel_grid = VtkVoxelGrid() vtk_voxel_grid.set_voxels(vx_grid_3d) # Create Voxel Grid Renderer in bottom half vtk_renderer = vtk.vtkRenderer() vtk_renderer.AddActor(vtk_boxes.vtk_actor) # vtk_renderer.AddActor(vtk_point_cloud.vtk_actor) vtk_renderer.AddActor(vtk_voxel_grid.vtk_actor) vtk_renderer.AddActor(axes) vtk_renderer.SetBackground(0.2, 0.3, 0.4) # Setup Camera current_cam = vtk_renderer.GetActiveCamera() current_cam.Pitch(170.0) current_cam.Roll(180.0) # Zooms out to fit all points on screen vtk_renderer.ResetCamera() # Zoom in slightly current_cam.Zoom(2.5) # Reset the clipping range to show all points vtk_renderer.ResetCameraClippingRange() # Setup Render Window vtk_render_window = vtk.vtkRenderWindow() vtk_render_window.SetWindowName("Anchors") vtk_render_window.SetSize(900, 500) vtk_render_window.AddRenderer(vtk_renderer) # Setup custom interactor style, which handles mouse and key events vtk_render_window_interactor = vtk.vtkRenderWindowInteractor() vtk_render_window_interactor.SetRenderWindow(vtk_render_window) vtk_render_window_interactor.SetInteractorStyle( vtk.vtkInteractorStyleTrackballCamera()) # Render in VTK vtk_render_window.Render() vtk_render_window_interactor.Start()
def _calculate_anchors_info(self, all_anchor_boxes_3d, empty_anchor_filter, gt_labels, calib_p2, image_shape): """Calculates the list of anchor information in the format: N x 14 [index, max_gt_iou, (6 x offsets), class_index, max_gt_img_iou, (4 x image offsets), image_class_index] max_gt_out - highest 3D iou with any ground truth box offsets - encoded offsets [dx, dy, dz, d_dimx, d_dimy, d_dimz] class_index - the anchor's class as an index (e.g. 0 or 1, for "Background" or "Car") max_gt_img_iou - highest image iou with any ground truth box image_offsets: encoded offsets [dx, dy, d_dimx, d_dimy] image_class_index: the anchor's class on image as an index (e.g. 0 or 1, for "Background" or "Car") calib_p2: stereo camera calibration p2 matrix image_shape: dimensions of the image [h, w] Args: all_anchor_boxes_3d: list of anchors in box_3d format N x [x, y, z, l, w, h, ry] empty_anchor_filter: boolean mask of which anchors are non empty gt_labels: list of Object Label data format containing ground truth labels to generate positives/negatives from. Returns: list of anchor info """ # Check for ground truth objects if len(gt_labels) == 0: raise Warning("No valid ground truth label to generate anchors.") kitti_utils = self._dataset.kitti_utils # Filter empty anchors anchor_indices = np.where(empty_anchor_filter)[0] anchor_boxes_3d = all_anchor_boxes_3d[empty_anchor_filter] # Convert anchor_boxes_3d to anchor format anchors = box_3d_encoder.box_3d_to_anchor(anchor_boxes_3d) # Convert gt to boxes_3d -> anchors -> iou format gt_boxes_3d = np.asarray([ box_3d_encoder.object_label_to_box_3d(gt_obj) for gt_obj in gt_labels ]) gt_anchors = box_3d_encoder.box_3d_to_anchor(gt_boxes_3d, ortho_rotate=True) rpn_iou_type = self.mini_batch_utils.rpn_iou_type if rpn_iou_type == '2d': # Convert anchors to 2d iou format anchors_for_2d_iou, _ = np.asarray( anchor_projector.project_to_bev(anchors, kitti_utils.bev_extents)) gt_boxes_for_2d_iou, _ = anchor_projector.project_to_bev( gt_anchors, kitti_utils.bev_extents) elif rpn_iou_type == '3d': # Convert anchors to 3d iou format for calculation anchors_for_3d_iou = box_3d_encoder.box_3d_to_3d_iou_format( anchor_boxes_3d) gt_boxes_for_3d_iou = \ box_3d_encoder.box_3d_to_3d_iou_format(gt_boxes_3d) else: raise ValueError('Invalid rpn_iou_type {}', rpn_iou_type) anchors_on_img, _ = np.asarray( anchor_projector.project_to_image_space(anchors, calib_p2, image_shape)) gt_img_boxes_for_2d_iou = np.asarray( [[gt_obj.x1, gt_obj.y1, gt_obj.x2, gt_obj.y2] for gt_obj in gt_labels]) # Initialize sample and offset lists num_anchors = len(anchor_boxes_3d) all_info = np.zeros((num_anchors, self.mini_batch_utils.col_length)) # Update anchor indices all_info[:, self.mini_batch_utils.col_anchor_indices] = anchor_indices # For each of the labels, generate samples for gt_idx in range(len(gt_labels)): gt_obj = gt_labels[gt_idx] gt_box_3d = gt_boxes_3d[gt_idx] # Get 2D or 3D IoU for every anchor if self.mini_batch_utils.rpn_iou_type == '2d': gt_box_for_2d_iou = gt_boxes_for_2d_iou[gt_idx] ious = evaluation.two_d_iou(gt_box_for_2d_iou, anchors_for_2d_iou) elif self.mini_batch_utils.rpn_iou_type == '3d': gt_box_for_3d_iou = gt_boxes_for_3d_iou[gt_idx] ious = evaluation.three_d_iou(gt_box_for_3d_iou, anchors_for_3d_iou) # Only update indices with a higher iou than before update_indices = np.greater( ious, all_info[:, self.mini_batch_utils.col_ious]) # Get ious to update ious_to_update = ious[update_indices] # Calculate offsets, use 3D iou to get highest iou anchors_to_update = anchors[update_indices] gt_anchor = box_3d_encoder.box_3d_to_anchor(gt_box_3d, ortho_rotate=True) offsets = anchor_encoder.anchor_to_offset(anchors_to_update, gt_anchor) # Convert gt type to index class_idx = kitti_utils.class_str_to_index(gt_obj.type) # Update anchors info (indices already updated) # [index, iou, (offsets), class_index] all_info[update_indices, self.mini_batch_utils.col_ious] = ious_to_update all_info[update_indices, self.mini_batch_utils.col_offsets_lo:self. mini_batch_utils.col_offsets_hi] = offsets all_info[update_indices, self.mini_batch_utils.col_class_idx] = class_idx # Image part gt_img_box_for_2d_iou = gt_img_boxes_for_2d_iou[gt_idx] img_ious = evaluation.two_d_iou(gt_img_box_for_2d_iou, anchors_on_img) # Only update indices with a higher iou than before update_img_indices = np.greater( img_ious, all_info[:, self.mini_batch_utils.col_img_ious]) # Get ious to update img_ious_to_update = img_ious[update_img_indices] #Calculate image offsets anchors_on_img_to_update = anchors_on_img[update_img_indices] img_offsets = anchor_encoder.img_box_to_offset( anchors_on_img_to_update, gt_img_box_for_2d_iou) # Update anchors info all_info[update_img_indices, self.mini_batch_utils.col_img_ious] = img_ious_to_update all_info[update_img_indices, self.mini_batch_utils.col_img_offsets_lo:self. mini_batch_utils.col_img_offsets_hi] = img_offsets all_info[update_img_indices, self.mini_batch_utils.col_img_class_idx] = class_idx return all_info
def preprocess(self, indices): """Preprocesses anchor info and saves info to files Args: indices (int array): sample indices to process. If None, processes all samples """ # Get anchor stride for class anchor_strides = self._anchor_strides dataset = self._dataset dataset_utils = self._dataset.kitti_utils classes_name = dataset.classes_name # Make folder if it doesn't exist yet output_dir = self.mini_batch_utils.get_file_path(classes_name, anchor_strides, sample_name=None) if not os.path.exists(output_dir): os.makedirs(output_dir) # Get clusters for class all_clusters_sizes, _ = dataset.get_cluster_info() anchor_generator = grid_anchor_3d_generator.GridAnchor3dGenerator() # Load indices of data_split all_samples = dataset.sample_list if indices is None: indices = np.arange(len(all_samples)) num_samples = len(indices) # For each image in the dataset, save info on the anchors for sample_idx in indices: # Get image name for given cluster sample_name = all_samples[sample_idx].name img_idx = int(sample_name) # Check for existing files and skip to the next if self._check_for_existing(classes_name, anchor_strides, sample_name): print("{} / {}: Sample already preprocessed".format( sample_idx + 1, num_samples, sample_name)) continue # Get ground truth and filter based on difficulty ground_truth_list = obj_utils.read_labels(dataset.label_dir, img_idx) # Get calibration matrix calib_p2 = calib_utils.read_calibration(self.calib_dir, int(sample_name)).p2 # Filter objects to dataset classes filtered_gt_list = dataset_utils.filter_labels(ground_truth_list) filtered_gt_list = np.asarray(filtered_gt_list) # Filtering by class has no valid ground truth, skip this image if len(filtered_gt_list) == 0: print("{} / {} No {}s for sample {} " "(Ground Truth Filter)".format(sample_idx + 1, num_samples, classes_name, sample_name)) # Output an empty file and move on to the next image. self._save_to_file(classes_name, anchor_strides, sample_name) continue # Get ground plane ground_plane = obj_utils.get_road_plane(img_idx, dataset.planes_dir) image = Image.open(dataset.get_rgb_image_path(sample_name)) image_shape = [image.size[1], image.size[0]] # Generate sliced 2D voxel grid for filtering vx_grid_2d = dataset_utils.create_sliced_voxel_grid_2d( sample_name, source=dataset.bev_source, image_shape=image_shape) # List for merging all anchors all_anchor_boxes_3d = [] # Create anchors for each class for class_idx in range(len(dataset.classes)): # Generate anchors for all classes grid_anchor_boxes_3d = anchor_generator.generate( area_3d=self._area_extents, anchor_3d_sizes=all_clusters_sizes[class_idx], anchor_stride=self._anchor_strides[class_idx], ground_plane=ground_plane) all_anchor_boxes_3d.extend(grid_anchor_boxes_3d) # Filter empty anchors all_anchor_boxes_3d = np.asarray(all_anchor_boxes_3d) anchors = box_3d_encoder.box_3d_to_anchor(all_anchor_boxes_3d) empty_anchor_filter = anchor_filter.get_empty_anchor_filter_2d( anchors, vx_grid_2d, self._density_threshold) # Calculate anchor info anchors_info = self._calculate_anchors_info( all_anchor_boxes_3d, empty_anchor_filter, filtered_gt_list, calib_p2, image_shape) anchor_ious = anchors_info[:, self.mini_batch_utils.col_ious] valid_iou_indices = np.where(anchor_ious > 0.0)[0] print("{} / {}:" "{:>6} anchors, " "{:>6} iou > 0.0, " "for {:>3} {}(s) for sample {}".format( sample_idx + 1, num_samples, len(anchors_info), len(valid_iou_indices), len(filtered_gt_list), classes_name, sample_name)) # Save anchors info self._save_to_file(classes_name, anchor_strides, sample_name, anchors_info)