def test_box_3d_to_anchor_projected(self):
"""
Check that boxes are projected with ortho_rotate=False,
and that projected boxes have the correct dimensions
"""
thetas = np.arange(0, 2 * np.pi, np.pi / 6)
boxes_3d = []
for theta in thetas:
boxes_3d.append([1, 2, 3, 4, 5, 6, theta])
boxes_3d = np.asarray(boxes_3d, dtype=np.float64)
cos_thetas = np.abs(np.cos(thetas))
sin_thetas = np.abs(np.sin(thetas))
expected_dims_x = 4 * cos_thetas + 5 * sin_thetas
expected_dims_z = 4 * sin_thetas + 5 * cos_thetas
expected_anchors = []
for exp_x, exp_z in zip(expected_dims_x, expected_dims_z):
expected_anchors.append([1, 2, 3, exp_x, 6, exp_z])
expected_anchors = np.asarray(expected_anchors, np.float64)
anchors = box_3d_encoder.box_3d_to_anchor(boxes_3d, ortho_rotate=False)
np.testing.assert_allclose(anchors, expected_anchors)
def test_box_3d_to_anchor_180_270(self):
box_3d = np.asarray([[1, 2, 3, 4, 5, 6, np.pi],
[1, 2, 3, 4, 5, 6, 3 * np.pi / 2]],
dtype=np.float64)
expected_anchors = np.asarray([[1, 2, 3, 4, 6, 5],
[1, 2, 3, 5, 6, 4]],
dtype=np.float64)
anchors = box_3d_encoder.box_3d_to_anchor(box_3d)
np.testing.assert_allclose(anchors, expected_anchors)
def _keep_near_anchors(self, all_anchor_boxes_3d,
gt_labels, dist_thres=10):
kitti_utils = self._dataset.kitti_utils
# Convert anchor_boxes_3d to anchor format
anchors = box_3d_encoder.box_3d_to_anchor(all_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)
keep = np.zeros(all_anchor_boxes_3d.shape[0], dtype=np.int32)
for gt_ind in range(gt_anchors.shape[0]):
gt_anchor = np.reshape(gt_anchors[gt_ind], (1, 6))
dist = np.linalg.norm(anchors[:, :3] - gt_anchor[:, :3], axis=1)
keep[dist < dist_thres] = 1
return keep.astype(np.bool)
def test_box_3d_to_anchor(self):
# box_3d format is [x, y, z, l, w, h, ry]
box_3d = np.asarray([[1, 2, 3, 4, 5, 6, 0],
[0, 0, 0, 1, 2, 3, 0],
[0, 0, 0, 1, 2, 3, np.pi / 2]],
dtype=np.float64)
# anchor format is [x, y, z, dim_x, dim_y, dim_z]
expected_anchors = np.asarray([[1, 2, 3, 4, 6, 5],
[0, 0, 0, 1, 3, 2],
[0, 0, 0, 2, 3, 1]],
dtype=np.float64)
anchors = box_3d_encoder.box_3d_to_anchor(box_3d)
np.testing.assert_allclose(anchors, expected_anchors)
def test_box_3d_to_anchor_rotated(self):
"""
Check that rotated boxes are rotated to the nearest 90
and that the dimensions do not change
"""
# Boxes at ry = 144, 288 should give same results as ry = 180, 270
box_3d = np.asarray([[1, 2, 3, 4, 5, 6, np.pi * 4 / 5],
[1, 2, 3, 4, 5, 6, 8 * np.pi / 5]],
dtype=np.float64)
expected_anchors = np.asarray([[1, 2, 3, 4, 6, 5],
[1, 2, 3, 5, 6, 4]],
dtype=np.float64)
anchors = box_3d_encoder.box_3d_to_anchor(box_3d, ortho_rotate=True)
np.testing.assert_allclose(anchors, expected_anchors)
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)
label_score_2d_path = self.get_score_2d_path(sample_name)
with open(label_score_2d_path, 'rb') as file:
label_score_2d = np.load(file)
file.close()
#label_score_2d = np.zeros((12, 39, 2))
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_classes = np.zeros(1)
label_score_2d = np.zeros((12, 39, 2))
img_idx = int(sample_name)
# 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]
image_input = rgb_image
cv_bgr_image_r = cv2.imread(self.get_rgb_image_r_path(sample_name))
rgb_image_r = cv_bgr_image_r[..., ::-1]
image_r_input = rgb_image_r
image_r_shape = rgb_image_r.shape[0:2]
# Get ground plane
ground_plane = obj_utils.get_road_plane(int(sample_name),
self.planes_dir)
# Get calibration
stereo_calib_p2 = calib_utils.read_calibration(
self.calib_dir, int(sample_name)).p2
stereo_calib_p3 = calib_utils.read_calibration(
self.calib_dir, int(sample_name)).p3
"""
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:
image_input = kitti_aug.flip_image(image_input)
image_r_input = kitti_aug.flip_image(image_r_input)
image_input, image_r_input = image_r_input, image_input
#ipoint_cloud = kitti_aug.flip_point_cloud(point_cloud)
obj_labels = [kitti_aug.flip_label_in_3d_only(obj)
for obj in obj_labels]
ground_plane = kitti_aug.flip_ground_plane(ground_plane)
stereo_calib_p2 = kitti_aug.flip_stereo_calib_p2(
stereo_calib_p2, image_shape)
stereo_calib_p3 = kitti_aug.flip_stereo_calib_p2(
stereo_calib_p3, image_r_shape)
stereo_calib_p2, stereo_calib_p3 = stereo_calib_p3, stereo_calib_p2
"""
# Augmentation (Image Jitter)
if kitti_aug.AUG_PCA_JITTER in sample.augs:
#print('Jittering')
image_input[:, :,
0:3] = kitti_aug.apply_pca_jitter(image_input[:, :,
0:3])
image_r_input[:, :, 0:3] = kitti_aug.apply_pca_jitter(
image_r_input[:, :, 0:3])
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_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)
else:
label_anchors = np.zeros((1, 6))
label_boxes_3d = np.zeros((1, 7))
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')
density_map = bev_images.get('density_map')
bev_input = np.dstack((*height_maps, density_map))
"""
sample_dict = {
constants.KEY_LABEL_BOXES_3D:
label_boxes_3d,
constants.KEY_LABEL_ANCHORS:
label_anchors,
constants.KEY_LABEL_CLASSES:
label_classes,
constants.KEY_LABEL_SCORE_2D:
label_score_2d,
constants.KEY_IMAGE_INPUT:
image_input,
constants.KEY_IMAGE_R_INPUT:
image_r_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_STEREO_CALIB_P3:
stereo_calib_p3,
constants.KEY_SAMPLE_NAME:
sample_name,
constants.KEY_SAMPLE_AUGS:
sample.augs
}
sample_dicts.append(sample_dict)
return sample_dicts
def main():
"""This demo shows RPN proposals and AVOD 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
avod_score_threshold = 0.1
# gt_classes = ['Car']
gt_classes = ['Pedestrian', 'Cyclist']
# gt_classes = ['Car', 'Pedestrian', 'Cyclist']
# Overwrite this to select a specific checkpoint
global_step = None
checkpoint_name = sys.argv[1] #'pyramid_cars_with_aug_example'
# 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 = True
# Save images for samples with no detections
save_empty_images = True
draw_score = True
draw_iou = True
##############################
# End of Options
##############################
# Get the dataset
dataset = DatasetBuilder.build_kitti_dataset(dataset_config)
# Setup Paths
predictions_dir = avod.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, avod_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, avod_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
avod_score_mask = prediction_scores >= avod_score_threshold
prediction_boxes_3d = prediction_boxes_3d[avod_score_mask]
prediction_scores = prediction_scores[avod_score_mask]
prediction_class_indices = \
prediction_class_indices[avod_score_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
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)
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)
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)
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 _fill_anchor_pl_inputs(self, anchors_info, ground_plane, image_shape,
stereo_calib_p2, sample_name, sample_augs):
"""
Fills anchor placeholder inputs with corresponding data
Args:
anchors_info: anchor info from mini_batch_utils
ground_plane: ground plane coefficients
image_shape: image shape (h, w), used for projecting anchors
sample_name: name of the sample, e.g. "000001"
sample_augs: list of sample augmentations
"""
# Lists for merging anchors info
all_anchor_boxes_3d = []
anchors_ious = []
anchor_offsets = []
anchor_classes = []
# Create anchors for each class
if len(self.dataset.classes) > 1:
for class_idx in range(len(self.dataset.classes)):
# Generate anchors for all classes
grid_anchor_boxes_3d = self._anchor_generator.generate(
area_3d=self._area_extents,
anchor_3d_sizes=self._cluster_sizes[class_idx],
anchor_stride=self._anchor_strides[class_idx],
ground_plane=ground_plane)
all_anchor_boxes_3d.append(grid_anchor_boxes_3d)
all_anchor_boxes_3d = np.concatenate(all_anchor_boxes_3d)
else:
# Don't loop for a single class
class_idx = 0
grid_anchor_boxes_3d = self._anchor_generator.generate(
area_3d=self._area_extents,
anchor_3d_sizes=self._cluster_sizes[class_idx],
anchor_stride=self._anchor_strides[class_idx],
ground_plane=ground_plane)
all_anchor_boxes_3d = grid_anchor_boxes_3d
# Filter empty anchors
# Skip if anchors_info is []
sample_has_labels = True
if self._train_val_test in ['train', 'val']:
# Read in anchor info during training / validation
if anchors_info:
anchor_indices, anchors_ious, anchor_offsets, \
anchor_classes = anchors_info
anchor_boxes_3d_to_use = all_anchor_boxes_3d[anchor_indices]
else:
train_cond = (self._train_val_test == "train"
and self._train_on_all_samples)
eval_cond = (self._train_val_test == "val"
and self._eval_all_samples)
if train_cond or eval_cond:
sample_has_labels = False
else:
sample_has_labels = False
if not sample_has_labels:
# During testing, or validation with no anchor info, manually
# filter empty anchors
# TODO: share voxel_grid_2d with BEV generation if possible
voxel_grid_2d = \
self.dataset.kitti_utils.create_sliced_voxel_grid_2d(
sample_name, self.dataset.bev_source,
image_shape=image_shape)
# Convert to anchors and filter
anchors_to_use = box_3d_encoder.box_3d_to_anchor(
all_anchor_boxes_3d)
empty_filter = anchor_filter.get_empty_anchor_filter_2d(
anchors_to_use, voxel_grid_2d, density_threshold=1)
anchor_boxes_3d_to_use = all_anchor_boxes_3d[empty_filter]
# Convert lists to ndarrays
anchor_boxes_3d_to_use = np.asarray(anchor_boxes_3d_to_use)
anchors_ious = np.asarray(anchors_ious)
anchor_offsets = np.asarray(anchor_offsets)
anchor_classes = np.asarray(anchor_classes)
# Flip anchors and centroid x offsets for augmented samples
if kitti_aug.AUG_FLIPPING in sample_augs:
anchor_boxes_3d_to_use = kitti_aug.flip_boxes_3d(
anchor_boxes_3d_to_use, flip_ry=False)
if anchors_info:
anchor_offsets[:, 0] = -anchor_offsets[:, 0]
# Convert to anchors
anchors_to_use = box_3d_encoder.box_3d_to_anchor(
anchor_boxes_3d_to_use)
num_anchors = len(anchors_to_use)
# Project anchors into bev
bev_anchors, bev_anchors_norm = anchor_projector.project_to_bev(
anchors_to_use, self._bev_extents)
# Project box_3d anchors into image space
img_anchors, img_anchors_norm = \
anchor_projector.project_to_image_space(
anchors_to_use, stereo_calib_p2, image_shape)
# Reorder into [y1, x1, y2, x2] for tf.crop_and_resize op
self._bev_anchors_norm = bev_anchors_norm[:, [1, 0, 3, 2]]
self._img_anchors_norm = img_anchors_norm[:, [1, 0, 3, 2]]
# Fill in placeholder inputs
self._placeholder_inputs[self.PL_ANCHORS] = anchors_to_use
# If we are in train/validation mode, and the anchor infos
# are not empty, store them. Checking for just anchors_ious
# to be non-empty should be enough.
if self._train_val_test in ['train', 'val'] and \
len(anchors_ious) > 0:
self._placeholder_inputs[self.PL_ANCHOR_IOUS] = anchors_ious
self._placeholder_inputs[self.PL_ANCHOR_OFFSETS] = anchor_offsets
self._placeholder_inputs[self.PL_ANCHOR_CLASSES] = anchor_classes
# During test, or val when there is no anchor info
elif self._train_val_test in ['test'] or \
len(anchors_ious) == 0:
# During testing, or validation with no gt, fill these in with 0s
self._placeholder_inputs[self.PL_ANCHOR_IOUS] = \
np.zeros(num_anchors)
self._placeholder_inputs[self.PL_ANCHOR_OFFSETS] = \
np.zeros([num_anchors, 6])
self._placeholder_inputs[self.PL_ANCHOR_CLASSES] = \
np.zeros(num_anchors)
else:
raise ValueError(
'Got run mode {}, and non-empty anchor info'.format(
self._train_val_test))
self._placeholder_inputs[self.PL_BEV_ANCHORS] = bev_anchors
self._placeholder_inputs[self.PL_BEV_ANCHORS_NORM] = \
self._bev_anchors_norm
self._placeholder_inputs[self.PL_IMG_ANCHORS] = img_anchors
self._placeholder_inputs[self.PL_IMG_ANCHORS_NORM] = \
self._img_anchors_norm
def _calculate_anchors_info(self,
all_anchor_boxes_3d,
empty_anchor_filter,
gt_labels):
"""Calculates the list of anchor information in the format:
N x 8 [max_gt_2d_iou, max_gt_3d_iou, (6 x offsets), 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")
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)
# 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
return all_info
def load_samples_from_file(self, image_path, lidar_path, calib_dir):
""" 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 = []
sample = self.sample_list[0]
sample_name = sample.name
obj_labels = None
anchors_info = []
label_anchors = np.zeros((1, 6))
label_boxes_3d = np.zeros((1, 7))
label_classes = np.zeros(1)
# Load image (BGR -> RGB)
cv_bgr_image = cv2.imread(image_path)
rgb_image = cv_bgr_image[..., ::-1]
image_shape = rgb_image.shape[0:2]
image_input = rgb_image
# Get ground plane
ground_plane = obj_utils.get_road_plane_from_file(calib_dir)
# Get calibration
stereo_calib = calib_utils.read_raw_calibration(calib_dir)
stereo_calib_p2 = stereo_calib.p2
point_cloud = self.kitti_utils.get_point_cloud_from_file(
self.bev_source, stereo_calib, lidar_path, image_shape)
# Augmentation (Flipping)
if kitti_aug.AUG_FLIPPING in sample.augs:
image_input = kitti_aug.flip_image(image_input)
point_cloud = kitti_aug.flip_point_cloud(point_cloud)
obj_labels = [
kitti_aug.flip_label_in_3d_only(obj) for obj in obj_labels
]
ground_plane = kitti_aug.flip_ground_plane(ground_plane)
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:
image_input[:, :,
0:3] = kitti_aug.apply_pca_jitter(image_input[:, :,
0:3])
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_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)
else:
label_anchors = np.zeros((1, 6))
label_boxes_3d = np.zeros((1, 7))
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')
density_map = bev_images.get('density_map')
bev_input = np.dstack((*height_maps, density_map))
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: 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
}
sample_dicts.append(sample_dict)
return sample_dicts
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 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)
os.makedirs(output_dir, exist_ok=True)
# 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)
# 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)
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)
def load_samples(self,
indices,
sin_type=None,
sin_level=None,
sin_input_name=None,
gen_all_sin_inputs=False,
list_mask_2d=None):
""" 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 idx, sample_idx in enumerate(indices):
sample = self.sample_list[sample_idx]
sample_name = sample.name
if list_mask_2d:
mask_2d = list_mask_2d[idx]
else:
mask_2d = None
# 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_classes = np.zeros(1)
img_idx = int(sample_name)
# 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]
image_input = rgb_image
# Get ground plane
ground_plane = obj_utils.get_road_plane(int(sample_name),
self.planes_dir)
# Get calibration
stereo_calib_p2 = calib_utils.read_calibration(
self.calib_dir, int(sample_name)).p2
# Read lidar with subsampling (handled before other preprocessing)
if (sin_type == 'lowres') and (sin_input_name == 'lidar'):
stride_sub = get_stride_sub(sin_level)
point_cloud = get_point_cloud_sub(img_idx, self.calib_dir,
self.velo_dir, image_shape,
stride_sub)
elif (sin_type == 'lowres') and gen_all_sin_inputs:
stride_sub = get_stride_sub(sin_level)
point_cloud = get_point_cloud_sub(img_idx, self.calib_dir,
self.velo_dir, image_shape,
stride_sub)
else:
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:
image_input = kitti_aug.flip_image(image_input)
point_cloud = kitti_aug.flip_point_cloud(point_cloud)
obj_labels = [
kitti_aug.flip_label_in_3d_only(obj) for obj in obj_labels
]
ground_plane = kitti_aug.flip_ground_plane(ground_plane)
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:
image_input[:, :,
0:3] = kitti_aug.apply_pca_jitter(image_input[:, :,
0:3])
# Add Single Input Noise
if (sin_input_name in SINFields.SIN_INPUT_NAMES) and (
sin_type in SINFields.VALID_SIN_TYPES):
image_input, point_cloud = genSINtoInputs(
image_input,
point_cloud,
sin_type=sin_type,
sin_level=sin_level,
sin_input_name=sin_input_name,
mask_2d=mask_2d,
frame_calib_p2=stereo_calib_p2)
# Add Input Noise to all
if gen_all_sin_inputs:
image_input, point_cloud = genSINtoAllInputs(
image_input,
point_cloud,
sin_type=sin_type,
sin_level=sin_level,
mask_2d=mask_2d,
frame_calib_p2=stereo_calib_p2)
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_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)
else:
label_anchors = np.zeros((1, 6))
label_boxes_3d = np.zeros((1, 7))
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')
density_map = bev_images.get('density_map')
bev_input = np.dstack((*height_maps, density_map))
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: 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
}
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_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():
"""
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
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)
# 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=True)
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_classes = np.zeros(1)
img_idx = int(sample_name)
# 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]
image_input = rgb_image
# Get ground plane
ground_plane = obj_utils.get_road_plane(int(sample_name),
self.planes_dir)
# Get calibration
stereo_calib = calib_utils.read_calibration(
self.calib_dir, int(sample_name))
stereo_calib_p2 = stereo_calib.p2
point_cloud = self.kitti_utils.get_point_cloud(
self.bev_source, img_idx, image_shape)
# Augmentation (Flipping)
# WZN: the flipping augmentation flips both image(in camera frame), pointcloud (in Lidar frame), and calibration
#matrix(between cam and Lidar) so the correspondence is still true.
if kitti_aug.AUG_FLIPPING in sample.augs:
image_input = kitti_aug.flip_image(image_input)
point_cloud = kitti_aug.flip_point_cloud(point_cloud)
obj_labels = [
kitti_aug.flip_label_in_3d_only(obj) for obj in obj_labels
]
ground_plane = kitti_aug.flip_ground_plane(ground_plane)
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:
image_input[:, :,
0:3] = kitti_aug.apply_pca_jitter(image_input[:, :,
0:3])
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_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)
else:
label_anchors = np.zeros((1, 6))
label_boxes_3d = np.zeros((1, 7))
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, output_indices=self.output_indices)
#WZN produce input for sparse pooling
if self.output_indices:
voxel_indices = bev_images[1]
pts_in_voxel = bev_images[2]
bev_images = bev_images[0]
height_maps = bev_images.get('height_maps')
density_map = bev_images.get('density_map')
bev_input = np.dstack((*height_maps, density_map))
#import pdb
#pdb.set_trace()
#WZN produce input for sparse pooling
if self.output_indices:
sparse_pooling_input1 = produce_sparse_pooling_input(
gen_sparse_pooling_input_avod(
pts_in_voxel, voxel_indices, stereo_calib,
[image_shape[1], image_shape[0]],
bev_input.shape[0:2]),
stride=[1, 1])
#WZN: Note here avod padded the vgg input by 4, so add it
bev_input_padded = np.copy(bev_input.shape[0:2])
bev_input_padded[0] = bev_input_padded[0] + 4
sparse_pooling_input2 = produce_sparse_pooling_input(
gen_sparse_pooling_input_avod(
pts_in_voxel, voxel_indices, stereo_calib,
[image_shape[1], image_shape[0]], bev_input_padded),
stride=[8, 8])
sparse_pooling_input = [
sparse_pooling_input1, sparse_pooling_input2
]
else:
sparse_pooling_input = None
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: image_input,
constants.KEY_BEV_INPUT: bev_input,
#WZN: for sparse pooling
constants.KEY_SPARSE_POOLING_INPUT: sparse_pooling_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
}
sample_dicts.append(sample_dict)
return sample_dicts
