def generate_semantic_mask(self, split):
output_dir = os.path.join(self.kitti_path, 'training/semantic_mask')
with open(os.path.join(self.kitti_path, split + '.txt')) as f:
frame_ids = [line.rstrip('\n') for line in f]
depth_map_dir = os.path.join(self.kitti_path, 'training/depth_map_dense')
if not os.path.isdir(output_dir):
os.mkdir(output_dir)
for fid in frame_ids:
fid = int(fid)
depth_map = np.load(os.path.join(depth_map_dir, '{:06}.npy'.format(fid)))
pc_dense = self.to_point_cloud(depth_map, fid)
objects = self.kitti_dataset.get_label_objects(fid)
image = self.kitti_dataset.get_image(fid)
calib = self.kitti_dataset.get_calibration(fid) # 3 by 4 matrix
objects = filter(lambda obj: obj.type in type_whitelist and obj.difficulty in difficulties_whitelist, objects)
pts_mask = {
'NonObject': np.zeros((len(pc_dense),), dtype=np.bool),
'Car': np.zeros((len(pc_dense),), dtype=np.bool),
'Pedestrian': np.zeros((len(pc_dense),), dtype=np.bool),
'Cyclist': np.zeros((len(pc_dense),), dtype=np.bool)
}
for obj in objects:
_,obj_box_3d = utils.compute_box_3d(obj, calib.P)
_,mask = extract_pc_in_box3d(pc_dense, obj_box_3d)
pts_mask[obj.type] = np.logical_or(mask, pts_mask[obj.type])
pts_mask['NonObject'] = (pts_mask['Car'] + pts_mask['Pedestrian'] + pts_mask['Cyclist']) == 0
semantic_mask = np.ones(image.shape[0:2], dtype=np.int32) * 255
# disp_mask = np.ones(image.shape, dtype=np.int32) * 255
# color_map = {
# 'NonObject': (0,0,0),
# 'Car': (0,0,142),
# 'Pedestrian': (220, 20, 60),
# 'Cyclist': (20, 255, 60)
# }
for k,v in pts_mask.items():
pts_2d = calib.project_rect_to_image(pc_dense[v, :3])
for u, v in pts_2d.astype(np.int32):
if u >= image.shape[1] or v >= image.shape[0]:
continue
semantic_mask[v, u] = g_type2onehotclass[k]
# disp_mask[v, u] = color_map[k]
# cv2.imshow('semantic_mask', disp_mask/255.0)
# cv2.waitKey(0)
fout = os.path.join(output_dir, '{:06}.png'.format(fid))
# np.save(fout, semantic_mask)
cv2.imwrite(fout, semantic_mask)
print('save', fid)
def expand_points(self, pc_rect, proposal, calib, seed_ind, data_idx_str):
prop = copy.deepcopy(proposal)
prop.l += 1
prop.w += 1
prop_corners_image_2d, prop_corners_3d = utils.compute_box_3d(
prop, calib.P)
_, local_ind = extract_pc_in_box3d(pc_rect, prop_corners_3d)
local_points = pc_rect[local_ind]
local_seg = self.pc_seg[data_idx_str][local_ind]
fg_ind = local_seg == 1
keypoints = pc_rect[seed_ind]
# print(np.sum(seed_ind))
# print(np.sum(local_ind))
# print(np.sum(fg_ind))
print('before: ', keypoints.shape[0])
candidates_ind = np.logical_and(np.logical_not(seed_ind), local_ind)
candidates_ind = np.logical_and(candidates_ind,
self.pc_seg[data_idx_str] == 1)
candidates_points = pc_rect[candidates_ind]
print('candidates size: ', candidates_points.shape[0])
while True:
tree = KDTree(keypoints[:, :3], leaf_size=2)
new_kp = []
selected = np.zeros((len(candidates_points), ))
for i in range(len(candidates_points)):
if tree.query_radius(np.expand_dims(candidates_points[i, :3],
axis=0),
r=0.3,
count_only=True) >= 3:
new_kp.append(candidates_points[i])
selected[i] = 1
if len(new_kp) == 0:
break
keypoints = np.concatenate((keypoints, new_kp), axis=0)
candidates_points = candidates_points[selected == 0]
print('after: ', keypoints.shape[0])
return keypoints
def visualize(dataset,
frame_id,
prediction,
seg_mask=None,
show_3d=False,
output_dir=None):
fig_size = (12.42, 3.75)
is_video = type(dataset).__name__ == 'kitti_object_video'
# pred_fig, pred_2d_axes, pred_3d_axes = \
# vis_utils.visualization(dataset.image_dir,
# int(frame_id),
# display=False,
# fig_size=fig_size)
pred_fig, pred_3d_axes = vis_utils.visualize_single_plot(dataset.image_dir,
int(frame_id),
is_video,
flipped=False,
display=False,
fig_size=fig_size)
calib = dataset.get_calibration(frame_id) # 3 by 4 matrix
# 2d visualization
# draw groundtruth
# labels = dataset.get_label_objects(frame_id)
# labels = filter(lambda obj: obj.type in type_whitelist and obj.difficulty in difficulties_whitelist, labels)
# draw_boxes(labels, calib, pred_2d_axes)
# draw prediction on second image
pred_corners = draw_boxes(prediction, calib, pred_3d_axes)
if output_dir:
filename = os.path.join(output_dir,
'result_2d_image/%06d.png' % frame_id)
plt.savefig(filename)
plt.close(pred_fig)
else:
plt.show()
raw_input()
if seg_mask is not None:
img = dataset.get_image(frame_id)
img = cv2.resize(img, (1200, 360))
seg_img = add_mask_to_img(img, seg_mask)
filename = os.path.join(output_dir,
'result_seg_image/%06d.png' % frame_id)
cv2.imwrite(filename, seg_img)
if show_3d:
# 3d visualization
pc_velo = dataset.get_lidar(frame_id)
image = dataset.get_image(frame_id)
img_height, img_width = image.shape[0:2]
_, _, img_fov_inds = get_lidar_in_image_fov(pc_velo[:, 0:3], calib, 0,
0, img_width, img_height,
True)
pc_velo = pc_velo[img_fov_inds, :]
boxes3d_velo = []
mask = np.zeros((len(pc_velo), ))
for corners in pred_corners:
pts_velo = calib.project_rect_to_velo(corners)
boxes3d_velo.append(pts_velo)
_, obj_mask = extract_pc_in_box3d(pc_velo, pts_velo)
mask = np.logical_or(mask, obj_mask)
fig = draw_lidar(pc_velo, pts_color=(1, 1, 1))
fig = draw_lidar(pc_velo[mask == 1], fig=fig, pts_color=(1, 0, 0))
fig = draw_gt_boxes3d(boxes3d_velo,
fig,
draw_text=False,
color=(1, 1, 1))
# raw_input()
if output_dir:
filename = os.path.join(output_dir,
'result_3d_image/%06d.png' % frame_id)
mlab.savefig(filename, figure=fig)
def load_frame_data(self,
data_idx_str,
random_flip=False,
random_rotate=False,
random_shift=False,
pca_jitter=False):
'''load one frame'''
data_idx = int(data_idx_str)
# print(data_idx_str)
calib = self.kitti_dataset.get_calibration(data_idx) # 3 by 4 matrix
if self.is_training:
objects = self.kitti_dataset.get_label_objects(data_idx)
else:
objects = []
objects = filter(
lambda obj: obj.type in self.types_list and obj.difficulty in self.
difficulties_list, objects)
gt_boxes = [] # ground truth boxes
image = self.kitti_dataset.get_image(data_idx)
pc_velo = self.kitti_dataset.get_lidar(data_idx)
img_height, img_width = image.shape[0:2]
# data augmentation
if pca_jitter:
image = apply_pca_jitter(image)[0]
# if random_flip and np.random.random()>0.5: # 50% chance flipping
if random_flip and False:
pc_velo[:, 1] *= -1
# FIXME: to flip the scene, optical center also need to be adjusted
# dont use with image now
calib.P[0, 2] = img_width - calib.P[0, 2]
image = np.flip(image, axis=1)
for obj in objects:
obj.t = [-obj.t[0], obj.t[1], obj.t[2]]
# ensure that ry is [-pi, pi]
if obj.ry >= 0:
obj.ry = np.pi - obj.ry
else:
obj.ry = -np.pi - obj.ry
# use original point cloud for rpn
if not self.train_img_seg:
_, pc_image_coord, img_fov_inds = get_lidar_in_image_fov(
pc_velo[:, 0:3], calib, 0, 0, img_width, img_height, True)
pc_velo = pc_velo[img_fov_inds, :]
choice = np.random.choice(pc_velo.shape[0],
self.npoints,
replace=True)
point_set = pc_velo[choice, :]
pc_rect = np.zeros_like(point_set)
pc_rect[:, 0:3] = calib.project_velo_to_rect(point_set[:, 0:3])
pc_rect[:, 3] = point_set[:, 3]
else:
# use dense point cloud for training image segmentation
pc_dense = self.dense_points.load_dense_points(data_idx)
# get_lidar_in_image_fov
pts_2d = calib.project_rect_to_image(pc_dense)
fov_inds = (pts_2d[:,0]<img_width) & (pts_2d[:,0]>=0) & \
(pts_2d[:,1]<img_height) & (pts_2d[:,1]>=0)
# fov_inds = fov_inds & (pc_dense[:,2]>2.0)
pc_dense = pc_dense[fov_inds, :]
obj_mask = {
'Car': np.zeros((len(pc_dense), ), dtype=np.bool),
'Pedestrian': np.zeros((len(pc_dense), ), dtype=np.bool),
'Cyclist': np.zeros((len(pc_dense), ), dtype=np.bool)
}
for obj in objects:
_, obj_box_3d = utils.compute_box_3d(obj, calib.P)
_, mask = extract_pc_in_box3d(pc_dense, obj_box_3d)
obj_mask[obj.type] = np.logical_or(mask, obj_mask[obj.type])
bg_mask = (obj_mask['Car'] + obj_mask['Pedestrian'] +
obj_mask['Cyclist']) == 0
car = np.where(obj_mask['Car'])[0]
# oversampling
ped = np.where(obj_mask['Pedestrian'])[0]
if ped.shape[0] > 0:
#ped = ped[np.random.choice(ped.shape[0], car.shape[0], replace=True)]
ped = ped[np.random.choice(ped.shape[0],
ped.shape[0] * 5,
replace=True)]
cyc = np.where(obj_mask['Cyclist'])[0]
if cyc.shape[0] > 0:
#cyc = cyc[np.random.choice(cyc.shape[0], car.shape[0], replace=True)]
cyc = cyc[np.random.choice(cyc.shape[0],
cyc.shape[0] * 10,
replace=True)]
fg_num = car.shape[0] + ped.shape[0] + cyc.shape[0]
#print('total points: ', len(pc_dense))
#print('fg_num: {0}, car: {1}, ped: {2}, cyc: {3}'.format(fg_num, car.shape[0], ped.shape[0], cyc.shape[0]))
bg = np.where(bg_mask)[0]
bg = bg[np.random.choice(bg.shape[0],
max(self.npoints - fg_num, 0),
replace=True)]
choice = np.concatenate((car, ped, cyc, bg), axis=0)[:self.npoints]
#choice = np.random.choice(pc_dense.shape[0], self.npoints, replace=True)
pc_rect = np.zeros((self.npoints, 4))
pc_rect[:, :3] = pc_dense[choice, :]
#start = time.time()
seg_mask = np.zeros((pc_rect.shape[0]))
# data augmentation
# dont use with image now
if random_rotate and False:
ry = (np.random.random() - 0.5) * math.radians(
20) # -10~10 degrees
pc_rect[:, 0:3] = rotate_points_along_y(pc_rect[:, 0:3], ry)
for obj in objects:
obj.t = rotate_points_along_y(obj.t, ry)
obj.ry -= ry
# ensure that ry is [-pi, pi]
if obj.ry > np.pi:
obj.ry -= 2 * np.pi
elif obj.ry < -np.pi:
obj.ry += 2 * np.pi
proposal_of_point = {} # point index to proposal vector
gt_box_of_point = {} # point index to corners_3d
for obj in objects:
_, obj_box_3d = utils.compute_box_3d(obj, calib.P)
_, obj_mask = extract_pc_in_box3d(pc_rect, obj_box_3d)
if np.sum(obj_mask) == 0:
# label without 3d points
# print('skip object without points')
continue
seg_mask[obj_mask] = g_type2onehotclass[obj.type]
#seg_mask[obj_mask] = 1
gt_boxes.append(obj_box_3d)
obj_idxs = np.where(obj_mask)[0]
# data augmentation
# FIXME: jitter point will make valid loss growing
# Also may go out of image view
if random_shift and False: # jitter object points
pc_rect[obj_idxs, :3] = shift_point_cloud(
pc_rect[obj_idxs, :3], 0.02)
for idx in obj_idxs:
# to save time
if self.train_img_seg:
proposal_of_point[idx] = [0, [0, 0, 0], 0, 0, 0, [0, 0, 0]]
else:
proposal_of_point[idx] = box_encoder.encode(
obj, pc_rect[idx, :3])
gt_box_of_point[idx] = obj_box_3d
# self.viz_frame(pc_rect, seg_mask, gt_boxes)
# return pc_rect, seg_mask, proposal_of_point, gt_box_of_point, gt_boxes
calib_matrix = np.copy(calib.P)
calib_matrix[0, :] *= (1200.0 / image.shape[1])
calib_matrix[1, :] *= (360.0 / image.shape[0])
#print('construct', time.time() - start)
return {
'pointcloud': pc_rect,
'image': cv2.resize(image, (1200, 360)),
'calib': calib_matrix,
'mask_label': seg_mask,
'proposal_of_point': self.get_proposal_out(proposal_of_point),
'gt_box_of_point': self.get_gt_box_of_points(gt_box_of_point),
'gt_boxes': gt_boxes,
'pc_choice': choice
}
def get_one_sample(self, proposal, pc_rect, image, calib, iou, gt_box_3d,
gt_object, data_idx_str, img_seg_map):
'''convert to frustum sample format'''
prop_corners_image_2d, prop_corners_3d = utils.compute_box_3d(
proposal, calib.P)
if prop_corners_image_2d is None:
print('skip proposal behind camera')
return False
# get points within proposal box
# expand proposal
proposal_expand = copy.deepcopy(proposal)
proposal_expand.l += 0.5
proposal_expand.w += 0.5
proposal_expand.h += 0.5
_, prop_corners_3d = utils.compute_box_3d(proposal_expand, calib.P)
_, prop_inds = extract_pc_in_box3d(pc_rect, prop_corners_3d)
pc_in_prop_box = pc_rect[prop_inds, :]
seg_mask = np.zeros((pc_in_prop_box.shape[0]))
if len(pc_in_prop_box) == 0:
print('Reject proposal with no point')
return False
# Get frustum angle
box2d_center = calib.project_rect_to_image(np.array([proposal.t]))[0]
uvdepth = np.zeros((1, 3))
uvdepth[0, 0:2] = box2d_center
uvdepth[0, 2] = 20 # some random depth
box2d_center_rect = calib.project_image_to_rect(uvdepth)
frustum_angle = -1 * np.arctan2(box2d_center_rect[0, 2],
box2d_center_rect[0, 0])
if gt_object is not None:
obj_type = gt_object.type
_, inds = extract_pc_in_box3d(pc_in_prop_box, gt_box_3d)
seg_mask[inds] = 1
# Reject object with too few point
if np.sum(seg_mask) < 5:
print('Reject object with too few point')
return False
# Get 3D BOX heading
heading_angle = gt_object.ry
# Get 3D BOX size
box3d_size = np.array([gt_object.l, gt_object.w, gt_object.h])
else:
obj_type = 'NonObject'
gt_box_3d = np.zeros((8, 3))
heading_angle = 0
box3d_size = np.zeros((1, 3))
#frustum_angle = 0
rot_angle = self.get_center_view_rot_angle(frustum_angle)
# Get point cloud
if self.rotate_to_center:
point_set = self.get_center_view_point_set(pc_in_prop_box,
rot_angle)
else:
point_set = pc_in_prop_box
# ------------------------------ LABELS ----------------------------
# classification
# assert(obj_type in ['Car', 'Pedestrian', 'Cyclist', 'NonObject'])
assert (obj_type in type_whitelist)
cls_label = g_type2onehotclass[obj_type]
# Get center point of 3D box
if self.rotate_to_center:
box3d_center = self.get_center_view_box3d_center(
gt_box_3d, rot_angle)
else:
box3d_center = self.get_box3d_center(gt_box_3d)
# Heading
if self.rotate_to_center:
heading_angle = heading_angle - rot_angle
# Size
size_class, size_residual = size2class(box3d_size, obj_type)
angle_class, angle_residual = angle2class(heading_angle,
NUM_HEADING_BIN)
self.sample_id_counter += 1
return Sample(self.sample_id_counter, point_set, seg_mask, box3d_center, angle_class, angle_residual,\
size_class, size_residual, rot_angle, cls_label, proposal, heading_angle, iou, img_seg_map, calib.P)
def get_one_sample(self, proposal, pc_rect, image, calib, iou, gt_box_3d,
gt_object, data_idx_str, img_seg_map):
'''convert to frustum sample format'''
prop_corners_image_2d, prop_corners_3d = utils.compute_box_3d(
proposal, calib.P)
if prop_corners_image_2d is None:
print('skip proposal behind camera')
return False
# get points within proposal box
_, prop_inds = extract_pc_in_box3d(pc_rect, prop_corners_3d)
pc_in_prop_box = pc_rect[prop_inds, :]
seg_mask = np.zeros((pc_in_prop_box.shape[0]))
if len(pc_in_prop_box) == 0:
print('Reject proposal with no point')
return False
# Get frustum angle
image_points = calib.project_rect_to_image(pc_in_prop_box[:, :3])
expand_image_points = np.concatenate(
(prop_corners_image_2d, image_points), axis=0)
xmin, ymin = expand_image_points.min(0)
xmax, ymax = expand_image_points.max(0)
# TODO: frustum angle is important, make use of image
# use gt angle for testing
if gt_object is not None:
xmin, ymin, xmax, ymax = gt_object.box2d
box2d_center = np.array([(xmin + xmax) / 2.0, (ymin + ymax) / 2.0])
uvdepth = np.zeros((1, 3))
uvdepth[0, 0:2] = box2d_center
uvdepth[0, 2] = 20 # some random depth
box2d_center_rect = calib.project_image_to_rect(uvdepth)
frustum_angle = -1 * np.arctan2(box2d_center_rect[0, 2],
box2d_center_rect[0, 0])
if gt_object is not None:
obj_type = gt_object.type
# TODO: use dbscan instead of ground truth
#_,gt_inds = extract_pc_in_box3d(pc_rect, gt_box_3d)
#prop_inds = np.logical_or(prop_inds, gt_inds)
_, inds = extract_pc_in_box3d(pc_in_prop_box, gt_box_3d)
seg_mask[inds] = 1
# Reject object with too few point
if np.sum(seg_mask) < 5:
print('Reject object with too few point')
return False
# Get 3D BOX heading
heading_angle = gt_object.ry
# Get 3D BOX size
box3d_size = np.array([gt_object.l, gt_object.w, gt_object.h])
else:
obj_type = 'NonObject'
gt_box_3d = np.zeros((8, 3))
heading_angle = 0
box3d_size = np.zeros((1, 3))
#frustum_angle = 0
rot_angle = self.get_center_view_rot_angle(frustum_angle)
# Get point cloud
if self.rotate_to_center:
point_set = self.get_center_view_point_set(pc_in_prop_box,
rot_angle)
else:
point_set = pc_in_prop_box
# ------------------------------ LABELS ----------------------------
# classification
# assert(obj_type in ['Car', 'Pedestrian', 'Cyclist', 'NonObject'])
assert (obj_type in type_whitelist)
cls_label = g_type2onehotclass[obj_type]
# Get center point of 3D box
if self.rotate_to_center:
box3d_center = self.get_center_view_box3d_center(
gt_box_3d, rot_angle)
else:
box3d_center = self.get_box3d_center(gt_box_3d)
# Heading
if self.rotate_to_center:
heading_angle = heading_angle - rot_angle
# Size
size_class, size_residual = size2class(box3d_size, obj_type)
angle_class, angle_residual = angle2class(heading_angle,
NUM_HEADING_BIN)
self.sample_id_counter += 1
return Sample(self.sample_id_counter, point_set, seg_mask, box3d_center, angle_class, angle_residual,\
size_class, size_residual, rot_angle, cls_label, proposal, heading_angle, iou, img_seg_map, calib.P)
def get_sample(self, pc_rect, image, img_seg_map, calib, proposal_,
max_iou, max_idx, objects):
thres_low = 0.3
thres_high = 0.55
if max_iou >= thres_high:
label = objects[max_idx]
if max_iou < thres_low:
label = None
if self.is_training and max_iou >= thres_low and max_iou < thres_high:
return False
# expand proposal boxes
proposal_expand = copy.deepcopy(proposal_)
proposal_expand.l += 0.5
proposal_expand.w += 0.5
_, box_3d = utils.compute_box_3d(proposal_expand, calib.P)
_, mask = extract_pc_in_box3d(pc_rect, box_3d)
# ignore proposal with no points
if (np.sum(mask) == 0):
return False
points = pc_rect[mask, :]
points_with_feats = np.zeros((points.shape[0], self.num_channel))
points_with_feats[:, :6] = points # xyz, intensity, seg_one_hot
# pooled points canonical transformation
points_with_feats[:, :3] -= proposal_.t
points_with_feats[:, :3] = rotate_points_along_y(
points_with_feats[:, :3], proposal_.ry)
sample = {}
self.sample_id_counter += 1
sample['id'] = self.sample_id_counter
sample['class'] = 0
sample['pointcloud'] = points_with_feats
sample['image'] = cv2.resize(image, (1200, 360))
sample['img_seg_map'] = img_seg_map
sample['calib'] = np.copy(calib.P)
# scale projection matrix
sample['calib'][0, :] *= (1200.0 / image.shape[1])
sample['calib'][1, :] *= (360.0 / image.shape[0])
sample['proposal_box'] = np.array([
proposal_.t[0], proposal_.t[1], proposal_.t[2], proposal_.l,
proposal_.h, proposal_.w, proposal_.ry
])
sample['center_cls'] = np.zeros((2, ), dtype=np.int32)
sample['center_res'] = np.zeros((3, ))
sample['angle_cls'] = 0
sample['angle_res'] = 0
sample['size_cls'] = 0
sample['size_res'] = np.zeros((3, ))
sample['gt_box'] = np.zeros((8, 3))
sample['train_regression'] = max_iou >= thres_high
if label:
sample['class'] = g_type2onehotclass[label.type]
# rotation canonical transformation
label_norm = copy.deepcopy(label)
label_norm.ry = label.ry - proposal_.ry
obj_vec = box_encoder.encode(label_norm, proposal_.t)
sample['center_cls'] = obj_vec[0]
sample['center_res'] = obj_vec[1]
sample['angle_cls'] = obj_vec[2]
sample['angle_res'] = obj_vec[3]
sample['size_cls'] = obj_vec[4]
sample['size_res'] = obj_vec[5]
_, gt_box_3d = utils.compute_box_3d(label, calib.P)
sample['gt_box'] = gt_box_3d
#self.viz_sample(sample)
return sample
def load_frame_data(self, data_idx_str, aug):
'''load one frame'''
start = time.time()
data_idx = int(data_idx_str)
try:
with open(os.path.join(self.data_dir, data_idx_str + '.pkl'),
'rb') as fin:
rpn_out = pickle.load(fin)
# load image segmentation output
img_seg_map = np.load(
os.path.join(self.data_dir, data_idx_str + '_seg.npy'))
except Exception as e:
print(e)
return []
proposals = self.get_proposals(rpn_out)
#print(data_idx_str)
calib = self.kitti_dataset.get_calibration(data_idx) # 3 by 4 matrix
if self.is_training:
objects = self.kitti_dataset.get_label_objects(data_idx)
else:
# while testing, all proposals will have class 0
objects = []
image = self.kitti_dataset.get_image(data_idx)
pc_velo = self.kitti_dataset.get_lidar(data_idx)
img_height, img_width = image.shape[0:2]
_, pc_image_coord, img_fov_inds = get_lidar_in_image_fov(
pc_velo[:, 0:3], calib, 0, 0, img_width, img_height, True)
pc_velo = pc_velo[img_fov_inds, :]
# Same point sampling as RPN
choice = rpn_out['pc_choices']
point_set = pc_velo[choice, :]
# point_set = pc_velo
pc_rect = np.zeros_like(point_set)
pc_rect[:, 0:3] = calib.project_velo_to_rect(point_set[:, 0:3])
pc_rect[:, 3] = point_set[:, 3]
# Segmentation results from RPN
seg_one_hot = np.zeros((pc_rect.shape[0], 2))
bg_ind = rpn_out['segmentation'] == 0
fg_ind = rpn_out['segmentation'] == 1
seg_one_hot[bg_ind, 0] = 1
seg_one_hot[fg_ind, 1] = 1
pc_rect = np.concatenate((pc_rect, seg_one_hot), axis=-1) # 6 channels
objects = filter(
lambda obj: obj.type in self.types_list and obj.difficulty in self.
difficulties_list, objects)
gt_boxes = [] # ground truth boxes
gt_boxes_xy = []
recall = np.zeros((len(objects), ), dtype=np.int32)
for obj in objects:
_, obj_box_3d = utils.compute_box_3d(obj, calib.P)
# skip label with no point
_, obj_mask = extract_pc_in_box3d(pc_rect, obj_box_3d)
if np.sum(obj_mask) == 0:
continue
gt_boxes.append(obj_box_3d)
gt_boxes_xy.append(obj_box_3d[:4, [0, 2]])
positive_samples = []
negative_samples = []
show_boxes = []
# boxes_2d = []
def process_proposal(prop):
b2d, prop_box_3d = utils.compute_box_3d(prop, calib.P)
prop_box_xy = prop_box_3d[:4, [0, 2]]
max_idx, max_iou = find_match_label(prop_box_xy, gt_boxes_xy)
sample = self.get_sample(pc_rect, image, img_seg_map, calib, prop,
max_iou, max_idx, objects)
# print(max_iou)
if not sample:
return -1
if sample['class'] == 0:
show_boxes.append(prop_box_3d)
negative_samples.append(sample)
else:
positive_samples.append(sample)
show_boxes.append(prop_box_3d)
recall[max_idx] = 1
return sample['class']
aug_proposals = []
AUG_X = {1: 1, 2: 2, 3: 2}
for prop in proposals:
cls = process_proposal(prop)
if not aug or cls <= 0:
continue
for x in range(AUG_X[cls]):
prop_ = random_shift_box3d(copy.deepcopy(prop), 0.1)
aug_proposals.append(prop_)
# add more proposals using label to increase training samples
'''
if self.split == 'train':
miss_objs = [objects[i] for i in range(len(objects)) if recall[i]==0]
aug_proposals += self.fill_proposals_with_gt(miss_objs)
'''
for prop in aug_proposals:
process_proposal(prop)
if self.is_training:
random.shuffle(negative_samples)
samples = positive_samples + negative_samples[:len(positive_samples
)]
else:
samples = positive_samples + negative_samples
random.shuffle(samples)
# self.viz_frame(pc_rect, np.zeros((pc_rect.shape[0],)), gt_boxes, show_boxes)
return samples
def stat_proposal(self):
'''statistic of proposals'''
total_iou_3d = 0
total_iou_2d = 0
total_angle_res = 0
total = 0
total_num = 0
total_recall = 0
total_gt = 0
for frame_id in self.frame_ids:
print(frame_id)
try:
with open(os.path.join(self.data_dir, frame_id+'.pkl'), 'rb') as fin:
rpn_out = pickle.load(fin)
except Exception as e:
print(e)
continue
data_idx = int(frame_id)
pc_velo = self.kitti_dataset.get_lidar(data_idx)
image = self.kitti_dataset.get_image(data_idx)
calib = self.kitti_dataset.get_calibration(data_idx) # 3 by 4 matrix
img_height, img_width = image.shape[0:2]
_, pc_image_coord, img_fov_inds = get_lidar_in_image_fov(pc_velo[:,0:3],
calib, 0, 0, img_width, img_height, True)
pc_velo = pc_velo[img_fov_inds, :]
pc_rect = np.zeros_like(pc_velo)
pc_rect[:,0:3] = calib.project_velo_to_rect(pc_velo[:,0:3])
pc_rect[:,3] = pc_velo[:,3]
objects = self.kitti_dataset.get_label_objects(data_idx)
objects = filter(lambda obj: obj.type in self.types_list and obj.difficulty in self.difficulties_list, objects)
proposals = self.get_proposals(rpn_out)
gt_boxes = [] # ground truth boxes
gt_boxes_xy = []
recall = np.zeros((len(objects),))
for obj in objects:
_,obj_box_3d = utils.compute_box_3d(obj, calib.P)
# skip label with no point
_,obj_mask = extract_pc_in_box3d(pc_rect, obj_box_3d)
if np.sum(obj_mask) == 0:
continue
gt_boxes.append(obj_box_3d)
gt_boxes_xy.append(obj_box_3d[:4, [0,2]])
for prop in proposals:
b2d,prop_box_3d = utils.compute_box_3d(prop, calib.P)
prop_box_xy = prop_box_3d[:4, [0,2]]
gt_idx, gt_iou = find_match_label(prop_box_xy, gt_boxes_xy)
if gt_idx == -1:
continue
if gt_iou >= 0.5:
recall[gt_idx] = 1
total_iou_2d += gt_iou
ry_residual = abs(prop.ry - objects[gt_idx].ry)
total_angle_res += ry_residual
total += 1
total_recall += np.sum(recall)
total_gt += len(objects)
total_num += len(proposals)
print('Average IOU 2d {0}'.format(total_iou_2d/total))
print('Average angle residual {0}'.format(total_angle_res/total))
print('Average proposal number: {0}'.format(total_num/len(self.frame_ids)))
print('Recall: {0}'.format(float(total_recall)/total_gt))
def stat(kitti_path, split, data_dir, type_whitelist, difficulties_whitelist):
kitti_dataset = kitti_object(kitti_path, 'training')
with open(os.path.join(kitti_path, split + '.txt')) as f:
frame_ids = [line.rstrip('\n') for line in f]
missed_sample = 0
missed_seg = 0
missed_seg_sample = 0
missed_obj_point = 0
tp = 0
fp = 0
fn = 0
total_recall = {10: 0, 20: 0, 30: 0, 40: 0, 50: 0, 100: 0, 300: 0}
label_count = 0
for frame_id in frame_ids:
print(frame_id)
sys.stdout.flush()
try:
with open(os.path.join(data_dir, frame_id + '.pkl'), 'rb') as fin:
rpn_out = pickle.load(fin)
except Exception as e:
print(e)
continue
data_idx = int(frame_id)
pc_choices = rpn_out['pc_choices']
seg_pred_point = rpn_out['segmentation'] # point seg
seg_pred = rpn_out['segmentation_fuse'] # point+img seg
fg_indices = rpn_out['fg_indices'] # after sampling
seg_pred = seg_pred_point
pc_velo = kitti_dataset.get_lidar(data_idx)
image = kitti_dataset.get_image(data_idx)
calib = kitti_dataset.get_calibration(data_idx) # 3 by 4 matrix
img_height, img_width = image.shape[0:2]
_, pc_image_coord, img_fov_inds = get_lidar_in_image_fov(
pc_velo[:, 0:3], calib, 0, 0, img_width, img_height, True)
pc_velo = pc_velo[img_fov_inds, :3]
pc_velo_sampled = pc_velo[pc_choices]
pc_rect = calib.project_velo_to_rect(pc_velo)
pc_rect_sampled = calib.project_velo_to_rect(pc_velo_sampled) # 16384
pc_seg = pc_rect_sampled[seg_pred > 0] # all seg fg points
if pc_seg.shape[0] > 0:
pc_seg_sampled = pc_rect_sampled[
fg_indices] # sampled fg points 2048
else:
pc_seg_sampled = pc_seg
objects = kitti_dataset.get_label_objects(data_idx)
objects = filter(
lambda obj: obj.type in type_whitelist and obj.difficulty in
difficulties_whitelist, objects)
gt_boxes = [] # ground truth boxes
gt_boxes_xy = []
seg_mask = np.zeros((pc_rect_sampled.shape[0], ))
for obj in objects:
_, obj_box_3d = utils.compute_box_3d(obj, calib.P)
# skip label with no point
_, obj_mask = extract_pc_in_box3d(pc_rect, obj_box_3d)
if np.sum(obj_mask) == 0:
continue
gt_boxes.append(obj_box_3d)
gt_boxes_xy.append(obj_box_3d[:4, [0, 2]])
label_count += 1
_, obj_mask = extract_pc_in_box3d(pc_rect_sampled, obj_box_3d)
seg_mask[obj_mask] = 1
if np.sum(obj_mask) == 0:
missed_sample += 1
_, mask = extract_pc_in_box3d(pc_seg, obj_box_3d)
if np.sum(mask) == 0:
missed_seg += 1
missed_obj_point += np.sum(obj_mask)
_, mask = extract_pc_in_box3d(pc_seg_sampled, obj_box_3d)
if np.sum(mask) == 0:
missed_seg_sample += 1
tp += np.sum(np.logical_and(seg_pred == seg_mask, seg_mask != 0))
fp += np.sum(np.logical_and(seg_pred != seg_mask, seg_mask == 0))
fn += np.sum(np.logical_and(seg_pred != seg_mask, seg_mask != 0))
for max_keep in total_recall.keys():
recall = np.zeros((len(objects), ))
proposals = get_proposals(rpn_out, 0.7, max_keep)
for prop in proposals:
if np.sum(recall) == len(objects):
break
b2d, prop_box_3d = utils.compute_box_3d(prop, calib.P)
prop_box_xy = prop_box_3d[:4, [0, 2]]
gt_idx, gt_iou = find_match_label(prop_box_xy, gt_boxes_xy)
if gt_idx == -1:
continue
if gt_iou >= 0.5:
recall[gt_idx] = 1
#if objects[gt_idx].type != 'Car' and gt_iou >= 0.5:
# print(objects[gt_idx].type)
total_recall[max_keep] += np.sum(recall)
print('Missed sample ratio: ', float(missed_sample) / label_count)
print('Missed seg ratio: ', float(missed_seg) / label_count)
print('Missed object average point: ', missed_obj_point / missed_seg)
print('Missed seg sample ratio: ', float(missed_seg_sample) / label_count)
for max_keep in total_recall.keys():
total_recall[max_keep] = float(total_recall[max_keep]) / label_count
print('Recall on proposal num: ', total_recall)
print('Seg recall: ', float(tp) / (tp + fn))
print('Seg precision: ', float(tp) / (tp + fp))
sys.stdout.flush()
def load_frame_data(self,
data_idx_str,
random_flip=False,
random_rotate=False,
random_shift=False,
pca_jitter=False):
'''load one frame'''
if self.use_aug_scene:
data_idx = int(data_idx_str[2:])
else:
data_idx = int(data_idx_str)
# print(data_idx_str)
calib = self.kitti_dataset.get_calibration(data_idx) # 3 by 4 matrix
image = self.kitti_dataset.get_image(data_idx)
img_height, img_width = image.shape[0:2]
# data augmentation
if pca_jitter:
image = apply_pca_jitter(image)[0]
objects = []
if not self.use_aug_scene or data_idx_str[:2] == '00':
pc_velo = self.kitti_dataset.get_lidar(data_idx)
_, pc_image_coord, img_fov_inds = get_lidar_in_image_fov(
pc_velo[:, 0:3], calib, 0, 0, img_width, img_height, True)
pc_velo = pc_velo[img_fov_inds, :]
choice = np.random.choice(pc_velo.shape[0],
self.npoints,
replace=True)
point_set = pc_velo[choice, :]
pc_rect = np.zeros_like(point_set)
pc_rect[:, 0:3] = calib.project_velo_to_rect(point_set[:, 0:3])
pc_rect[:, 3] = point_set[:, 3]
if self.is_training:
objects = self.kitti_dataset.get_label_objects(data_idx)
else:
pc_rect = utils.load_velo_scan(
os.path.join(
self.kitti_path,
'aug_scene/rectified_data/{0}.bin'.format(data_idx_str)))
choice = np.random.choice(pc_rect.shape[0],
self.npoints,
replace=True)
pc_rect = pc_rect[choice, :]
if self.is_training:
objects = utils.read_label(
os.path.join(
self.kitti_path,
'aug_scene/aug_label/{0}.txt'.format(data_idx_str)))
objects = filter(
lambda obj: obj.type in self.types_list and obj.difficulty in self.
difficulties_list, objects)
gt_boxes = [] # ground truth boxes
#start = time.time()
seg_mask = np.zeros((pc_rect.shape[0]))
# data augmentation
if random_flip and np.random.random() > 0.5: # 50% chance flipping
pc_rect[:, 0] *= -1
for obj in objects:
obj.t = [-obj.t[0], obj.t[1], obj.t[2]]
# ensure that ry is [-pi, pi]
if obj.ry >= 0:
obj.ry = np.pi - obj.ry
else:
obj.ry = -np.pi - obj.ry
if random_rotate:
ry = (np.random.random() - 0.5) * math.radians(
20) # -10~10 degrees
pc_rect[:, 0:3] = rotate_points_along_y(pc_rect[:, 0:3], ry)
for obj in objects:
obj.t = rotate_points_along_y(obj.t, ry)
obj.ry -= ry
# ensure that ry is [-pi, pi]
if obj.ry > np.pi:
obj.ry -= 2 * np.pi
elif obj.ry < -np.pi:
obj.ry += 2 * np.pi
proposal_of_point = {} # point index to proposal vector
gt_box_of_point = {} # point index to corners_3d
for obj in objects:
_, obj_box_3d = utils.compute_box_3d(obj, calib.P)
_, obj_mask = extract_pc_in_box3d(pc_rect, obj_box_3d)
if np.sum(obj_mask) == 0:
# label without 3d points
# print('skip object without points')
continue
# IMPORTANT: this must match with NUM_SEG_CLASSES
#seg_mask[obj_mask] = g_type2onehotclass[obj.type]
seg_mask[obj_mask] = 1
gt_boxes.append(obj_box_3d)
obj_idxs = np.where(obj_mask)[0]
# data augmentation
# FIXME: jitter point will make valid loss growing
# Also may go out of image view
if random_shift and False: # jitter object points
pc_rect[obj_idxs, :3] = shift_point_cloud(
pc_rect[obj_idxs, :3], 0.02)
for idx in obj_idxs:
proposal_of_point[idx] = box_encoder.encode(
obj, pc_rect[idx, :3])
gt_box_of_point[idx] = obj_box_3d
# self.viz_frame(pc_rect, seg_mask, gt_boxes)
# return pc_rect, seg_mask, proposal_of_point, gt_box_of_point, gt_boxes
calib_matrix = np.copy(calib.P)
calib_matrix[0, :] *= (1200.0 / image.shape[1])
calib_matrix[1, :] *= (360.0 / image.shape[0])
#print('construct', time.time() - start)
return {
'pointcloud': pc_rect,
'image': cv2.resize(image, (1200, 360)),
'calib': calib_matrix,
'mask_label': seg_mask,
'proposal_of_point': self.get_proposal_out(proposal_of_point),
'gt_box_of_point': self.get_gt_box_of_points(gt_box_of_point),
'gt_boxes': gt_boxes,
'pc_choice': choice
}