def draw_3d_bbox(self,
frame,
points_camera,
cam_calib,
cam_pose,
cam_near_clip: float = 0.15,
line_color: tuple = (0, 255, 0),
line_width: int = 3,
corner_info: str = None):
projpoints = tu.get_3d_bbox_vertex(cam_calib, cam_pose, points_camera,
cam_near_clip)
for p1, p2 in projpoints:
cv2.line(frame, (int(p1[0]), int(p1[1])), (int(p2[0]), int(p2[1])),
line_color, line_width)
if corner_info is not None:
is_before = False
cp1 = tu.cameratoimage(points_camera[0:1], cam_calib)[0]
if cp1 is not None:
is_before = tu.is_before_clip_plane_camera(
points_camera[0:1], cam_near_clip)[0]
if is_before:
x1 = int(cp1[0])
y1 = int(cp1[1])
frame = self.draw_corner_info(frame, x1, y1, corner_info,
line_color)
return frame
def draw_3d_traj(frame,
points_hist,
cam_calib,
cam_pose,
line_color=(0, 255, 0)):
# Plot center history
for index, wt in enumerate(points_hist):
ct = tu.worldtocamera(wt, cam_pose)
pt = tu.cameratoimage(ct, cam_calib)
rgba = line_color + tuple(
[int(max(float(index) / len(points_hist), 0.5) * 255)])
cv2.circle(frame, (int(pt[0, 0]), int(pt[0, 1])),
3,
rgba,
thickness=-1)
return frame
def convert_track(data_dir, mode=None, adjust_center=True):
kitti = defaultdict(list)
img_dir = osp.join(data_dir, 'image_02')
label_dir = osp.join(data_dir, 'label_02')
cali_dir = osp.join(data_dir, 'calib')
oxt_dir = osp.join(data_dir, 'oxts')
if not osp.exists(img_dir):
print(f"Folder {img_dir} is not found")
return None
if not osp.exists(label_dir):
label_dir = None
vid_names = sorted(os.listdir(img_dir))
print(f"{data_dir} with {len(vid_names)} sequences")
for k, v in cats_mapping.items():
kitti['categories'].append(dict(id=v, name=k))
img_id = 0
global_track_id = 0
ann_id = 0
for vid_id, vid_name in enumerate(vid_names):
if mode == 'train':
if vid_name in val_sets:
continue
elif mode == 'val':
if vid_name not in val_sets:
continue
elif mode == 'mini':
if vid_name not in mini_sets:
continue
print("VID ID: {}".format(vid_id))
ind2id = dict()
trackid_maps = dict()
img_names = sorted([
f.path for f in os.scandir(osp.join(img_dir, vid_name))
if f.is_file() and f.name.endswith('png')
])
vid_info = dict(id=vid_id, name=vid_name, n_frames=len(img_names))
kitti['videos'].append(vid_info)
projection = ku.read_calib(cali_dir, vid_id)
for fr, img_name in enumerate(sorted(img_names)):
img = mmcv.imread(img_name)
fields = ku.read_oxts(oxt_dir, vid_id)
poses = [ku.KittiPoseParser(fields[i]) for i in range(len(fields))]
rotation = R.from_matrix(poses[fr].rotation).as_euler('xyz')
position = poses[fr].position - poses[0].position
pose_dict = dict(rotation=rotation.tolist(),
position=position.tolist())
height, width, _ = img.shape
index = fr
img_info = dict(file_name=img_name,
cali=projection.tolist(),
pose=pose_dict,
height=height,
width=width,
fov=60,
near_clip=0.15,
id=img_id,
video_id=vid_id,
index=index)
kitti['images'].append(img_info)
ind2id[index] = img_id
img_id += 1
if label_dir:
label_file = osp.join(label_dir, '{}.txt'.format(vid_name))
labels = mmcv.list_from_file(label_file)
for label in labels:
label = label.split()
cat = label[2]
if cat in ['DontCare']:
continue
# if cat not in kitti_cats.keys():
# continue
image_id = ind2id[int(label[0])]
if label[1] in trackid_maps.keys():
track_id = trackid_maps[label[1]]
else:
track_id = global_track_id
trackid_maps[label[1]] = track_id
global_track_id += 1
x1, y1, x2, y2 = float(label[6]), float(label[7]), float(
label[8]), float(label[9])
if adjust_center:
# KITTI GT uses the bottom of the car as center (x, 0, z).
# Prediction uses center of the bbox as center (x, y, z).
# So we align them to the bottom center as GT does
y_cen_adjust = float(label[10]) / 2.0
else:
y_cen_adjust = 0.0
center_2d = tu.cameratoimage(
np.array([[
float(label[13]),
float(label[14]) - y_cen_adjust,
float(label[15])
]]), projection).flatten().tolist()
ann = dict(id=ann_id,
image_id=image_id,
category_id=cats_mapping[kitti_cats[cat]],
instance_id=track_id,
alpha=float(label[5]),
roty=float(label[16]),
dimension=[
float(label[10]),
float(label[11]),
float(label[12])
],
translation=[
float(label[13]),
float(label[14]) - y_cen_adjust,
float(label[15])
],
is_occluded=int(label[4]),
is_truncated=float(label[3]),
center_2d=center_2d,
delta_2d=[
center_2d[0] - (x1 + x2) / 2.0,
center_2d[1] - (y1 + y2) / 2.0
],
bbox=[x1, y1, x2 - x1, y2 - y1],
area=(x2 - x1) * (y2 - y1),
iscrowd=False,
ignore=False,
segmentation=[[x1, y1, x1, y2, x2, y2, x2, y1]])
kitti['annotations'].append(ann)
ann_id += 1
return kitti
def convert_det(data_dir, mode=None, adjust_center=True):
kitti = defaultdict(list)
img_dir = osp.join(data_dir, 'image_2')
label_dir = osp.join(data_dir, 'label_2')
cali_dir = osp.join(data_dir, 'calib')
if not osp.exists(img_dir):
print(f"Folder {img_dir} is not found")
return None
if not osp.exists(label_dir):
label_dir = None
img_names = sorted(os.listdir(img_dir))
for k, v in cats_mapping.items():
kitti['categories'].append(dict(id=v, name=k))
pose_dict = dict(rotation=[0, 0, 0], position=[0, 0, 0])
ann_id = 0
vid_info = dict(id=0, name='', n_frames=len(img_names))
kitti['videos'].append(vid_info)
for img_id, img_name in enumerate(img_names):
if mode == 'train':
if osp.splitext(img_name)[0] in det_val_sets:
continue
elif mode == 'val':
if osp.splitext(img_name)[0] not in det_val_sets:
continue
print('DET ID: {}'.format(img_id))
img = mmcv.imread(osp.join(img_dir, img_name))
height, width, _ = img.shape
projection = ku.read_calib_det(cali_dir, img_id)
index = int(img_name.split('.')[0])
img_info = dict(file_name=osp.join(img_dir, img_name),
cali=projection.tolist(),
pose=pose_dict,
height=height,
width=width,
fov=60,
near_clip=0.15,
id=img_id,
video_id=0,
index=index)
kitti['images'].append(img_info)
if label_dir:
label_file = osp.join(label_dir,
'{}.txt'.format(img_name.split('.')[0]))
labels = mmcv.list_from_file(label_file)
track_id = 0
for label in labels:
label = label.split()
cat = label[0]
if cat in ['DontCare']:
continue
# if cat not in kitti_cats.keys():
# continue
x1, y1, x2, y2 = float(label[4]), float(label[5]), float(
label[6]), float(label[7])
if adjust_center:
# KITTI GT uses the bottom of the car as center (x, 0, z).
# Prediction uses center of the bbox as center (x, y, z).
# So we align them to the bottom center as GT does
y_cen_adjust = float(label[8]) / 2.0
else:
y_cen_adjust = 0.0
center_2d = tu.cameratoimage(
np.array([[
float(label[11]),
float(label[12]) - y_cen_adjust,
float(label[13])
]]), projection).flatten().tolist()
ann = dict(id=ann_id,
image_id=img_id,
instance_id=track_id,
category_id=cats_mapping[kitti_cats[cat]],
alpha=float(label[3]),
roty=float(label[14]),
dimension=[
float(label[8]),
float(label[9]),
float(label[10])
],
translation=[
float(label[11]),
float(label[12]) - y_cen_adjust,
float(label[13])
],
is_occluded=int(label[2]),
is_truncated=float(label[1]),
center_2d=center_2d,
delta_2d=[
center_2d[0] - (x1 + x2) / 2.0,
center_2d[1] - (y1 + y2) / 2.0
],
bbox=[x1, y1, x2 - x1, y2 - y1],
area=(x2 - x1) * (y2 - y1),
iscrowd=False,
ignore=False,
segmentation=[[x1, y1, x1, y2, x2, y2, x2, y1]])
kitti['annotations'].append(ann)
ann_id += 1
track_id += 1
return kitti
def convert_track(data_dir, subset: str):
gta_anno = defaultdict(list)
set_size = data_set_dict[subset]['amount']
subset_folder = data_set_dict[subset]['folder']
img_dir = os.path.join(data_dir, subset_folder, 'image')
label_dir = os.path.join(data_dir, subset_folder, 'label')
if not osp.exists(img_dir):
print(f"Folder {img_dir} is not found")
return None
if not os.path.exists(label_dir):
label_dir = None
vid_names = sorted([f.path for f in os.scandir(label_dir) if f.is_dir()])
# Uniformly sample videos
if set_size < len(vid_names):
vid_names = vid_names[::set_size]
# get information at boxes level. Collect dict. per box, not image.
print(f"{subset} with {len(vid_names)} sequences")
for k, v in cats_mapping.items():
gta_anno['categories'].append(dict(id=v, name=k))
img_id = 0
global_track_id = 0
ann_id = 0
for vid_id, vid_name in enumerate(vid_names):
print(f"VID {vid_id} ID: {vid_name}")
ind2id = dict()
trackid_maps = dict()
fr_names = sorted([
f.path for f in os.scandir(vid_name)
if f.is_file() and f.name.endswith('final.json')
])
if vid_name == osp.join(
data_dir,
'train/label/rec_10090618_snow_10h14m_x-493y-1796tox-1884y1790'
):
print('Bump!')
continue
vid_info = dict(id=vid_id, name=vid_name, n_frames=len(fr_names))
gta_anno['videos'].append(vid_info)
init_position = bu.load_json(fr_names[0])['extrinsics']['location']
for fr_idx, fr_name in enumerate(fr_names):
frame = bu.load_json(fr_name)
img_name = fr_name.replace('label', 'image').replace('json', 'jpg')
height = frame['resolution']['height']
width = frame['resolution']['width']
rot_angle = np.array(frame['extrinsics']['rotation'])
rot_matrix = tu.angle2rot(rot_angle)
gps_to_camera = tu.angle2rot(np.array([np.pi / 2, 0, 0]),
inverse=True)
rot_matrix = rot_matrix.dot(gps_to_camera)
rotation = R.from_matrix(rot_matrix).as_euler('xyz')
position = [
float(p_t) - float(p_0) for (
p_t,
p_0) in zip(frame['extrinsics']['location'], init_position)
]
pose_dict = dict(rotation=rotation.tolist(), position=position)
projection = np.array(frame['intrinsics']['cali'])
index = fr_idx
img_info = dict(file_name=img_name,
cali=projection.tolist(),
pose=pose_dict,
height=height,
width=width,
fov=60,
near_clip=0.15,
timestamp=frame['timestamp'],
id=img_id,
video_id=vid_id,
index=index)
gta_anno['images'].append(img_info)
ind2id[index] = img_id
img_id += 1
for label in frame['labels']:
cat = label['category']
if cat in ['DontCare']:
continue
image_id = ind2id[index]
if label['id'] in trackid_maps.keys():
track_id = trackid_maps[label['id']]
else:
track_id = global_track_id
trackid_maps[label['id']] = track_id
global_track_id += 1
x1, y1, x2, y2 = float(label['box2d']['x1']), float(
label['box2d']['y1']), float(label['box2d']['x2']), float(
label['box2d']['y2'])
location = bu.get_label_array([label], ['box3d', 'location'],
(0, 3)).astype(float)
center_2d = tu.cameratoimage(location,
projection).flatten().tolist()
ann = dict(id=ann_id,
image_id=image_id,
category_id=cats_mapping[gta_merge_maps[cat]],
instance_id=track_id,
alpha=float(label['box3d']['alpha']),
roty=float(label['box3d']['orientation']),
dimension=[
float(dim)
for dim in label['box3d']['dimension']
],
translation=[
float(loc) for loc in label['box3d']['location']
],
is_occluded=int(label['attributes']['occluded']),
is_truncated=int(label['attributes']['truncated']),
center_2d=center_2d,
delta_2d=[
center_2d[0] - (x1 + x2) / 2.0,
center_2d[1] - (y1 + y2) / 2.0
],
bbox=[x1, y1, x2 - x1, y2 - y1],
area=(x2 - x1) * (y2 - y1),
iscrowd=False,
ignore=label['attributes']['ignore'],
segmentation=[[x1, y1, x1, y2, x2, y2, x2, y1]])
gta_anno['annotations'].append(ann)
ann_id += 1
return gta_anno