def build_pred(self):
seq_data = []
if self.data_path == []:
print("Empty data")
return seq_data
for idx, file_path in enumerate(self.data_path):
frame_data = bh.load_json(file_path)
if idx % args.verbose_interval == 0 and idx != 0:
print('{} images.'.format(idx))
match_box = self.det_result[idx]
prediction_data = []
for i, box in enumerate(match_box):
prediction = bh.init_labels_format()
prediction['box2d']['x1'] = int(box[0])
prediction['box2d']['y1'] = int(box[1])
prediction['box2d']['x2'] = int(box[2])
prediction['box2d']['y2'] = int(box[3])
prediction['box2d']['confidence'] = float(box[6])
# 3D box projected center
prediction['box3d']['xc'] = int(box[4])
prediction['box3d']['yc'] = int(box[5])
del prediction['poly2d']
del prediction['attributes']['trafficLightColor']
del prediction['attributes']['areaType']
del prediction['attributes']['laneDirection']
del prediction['attributes']['laneStyle']
del prediction['attributes']['laneTypes']
prediction_data.append(prediction)
frame_data['prediction'] = prediction_data
print("Frame {}, GT: {} Boxes, PD: {} Boxes".format(
idx,
len(frame_data['labels']),
len(frame_data['prediction'])))
del frame_data['labels']
filename = os.path.join(
DATASET.PRED_PATH,
frame_data['videoName'],
os.path.basename(file_path))
if not os.path.exists(os.path.dirname(filename)):
os.mkdir(os.path.dirname(filename))
bh.dump_json(filename, frame_data)
seq_data.append(filename)
print("Saving {} frames...".format(idx))
return seq_data
# Update input json path and save path
method = method + setting_str
trk_path = '{ROOT}{SESS}_{EP}_{DT}_{PH}_set/{MT}_pd.json'.format(
**{
'ROOT': args.root,
'SESS': args.session,
'EP': args.epoch,
'DT': args.set,
'PH': args.split,
'MT': method
})
save_path = '{ROOT}{SESS}_{EP}_{DT}_{PH}_set/{MT}/data/'.format(
**{
'ROOT': args.root,
'SESS': args.session,
'EP': args.epoch,
'DT': args.set,
'PH': args.split,
'MT': method
})
# Load tracked results
trk_result = bh.load_json(trk_path)
print(len(trk_result), len(data_path))
for seq_idx, dpath in enumerate(data_path):
trk_seq = [n['hypotheses'] for n in trk_result[seq_idx]['frames']]
det_seq = bh.load_json(dpath)
convert_app(det_seq, trk_seq,
os.path.join(save_path, os.path.basename(dpath)))
# Load data
print("Load label file from path: {}".format(DATASET.PRED_PATH))
folders, data_path = load_label_path(DATASET.PRED_PATH)
for seq_idx, folder in enumerate(folders):
pkl_path = '{ROOT}{SESS}_{EP}_{DT}_{PH}_set/{SESS}_{EP}_{' \
'SQ}_bdd_roipool_output.pkl'.format(
**{'ROOT': args.root, 'SESS': args.session, 'EP': args.epoch,
'DT': args.set, 'PH': args.split, 'SQ': folder})
save_path = '{ROOT}{SESS}_{EP}_{DT}_{PH}_set/{SQ}_bdd_3d.json'.format(
**{'ROOT': args.root, 'SESS': args.session, 'EP': args.epoch,
'DT': args.set, 'PH': args.split, 'SQ': folder})
det_pred = load_single_frame_result(pkl_path)
hypos = []
for fr_idx, frame_path in enumerate(data_path[folder]):
det_placeholder = bh.load_json(frame_path)
hypo = convert_app(det_placeholder, det_pred[fr_idx])
hypos.append(hypo)
print("Saving updated tracking results with {} frames at {}...".format(
len(hypos), save_path))
bh.dump_json(save_path, hypos)
def gta_label(self):
# get information at boxes level. Collect dict. per box, not image.
file_list = load_label_path(os.path.join(DATASET.LABEL_PATH))
print("{} with {} sequences".format(args.split, len(file_list)))
data_label_list = []
re_pattern = re.compile('rec_(.{8})_(.+)_(.+)h(.+)m_(.+)')
# rec_10090911_clouds_21h53m_x-968y-1487tox2523y214
# ('10090911', 'clouds', '21', '53', 'x-968y-1487tox2523y214')
for fidx, filename in enumerate(file_list):
print(fidx, filename)
dataset = bh.load_json(filename)
data_label = {}
for fr_idx, frame in enumerate(dataset):
img_name = os.path.join(frame['dset_name'],
str(frame['timestamp']) + '_final.jpg')
num_boxes = len(frame['object'])
obj = frame['object']
data_label[img_name] = {}
data_label[img_name]['vid_name'] = os.path.join(
frame['dset_name'])
img_log = re_pattern.match(frame['dset_name'])
img_weather = img_log.group(2)
img_hour = img_log.group(3)
data_label[img_name]['weather'] = img_weather
data_label[img_name]['timeofday'] = bh.get_time_of_day(
int(img_hour))
data_label[img_name]['timestamp'] = frame['timestamp']
data_label[img_name]['fov'] = frame['camera']['fov']
data_label[img_name]['nearClip'] = frame['camera']['nearClip']
data_label[img_name]['pose'] = {
'rotation': [np.pi * angle / 180.0
for angle in frame['pose']['rotation']],
'position': [p_t - p_0 for (p_t, p_0) in \
zip(frame['pose']['position'],
dataset[0]['pose']['position'])]}
data_label[img_name]['img_name'] = img_name
data_label[img_name]['num_boxes'] = num_boxes
data_label[img_name]['pixel'] = [
obj[i]['n_pixel'] for i in range(num_boxes)
]
data_label[img_name]['class'] = [
DATASET.CLASS_PARSER[obj[i]['kitti']['type']]
for i in range(num_boxes)
]
data_label[img_name]['ignore'] = [
obj[i]['ignore'] for i in range(num_boxes)
]
data_label[img_name]['tracking_id'] = [
obj[i]['tracking_id'] for i in range(num_boxes)
]
data_label[img_name]['truncated'] = [
obj[i]['kitti']['truncated'] for i in range(num_boxes)
]
data_label[img_name]['occluded'] = [
obj[i]['kitti']['occluded'] for i in range(num_boxes)
]
data_label[img_name]['boxes'] = [
obj[i]['kitti']['bbox'] for i in range(num_boxes)
]
data_label[img_name]['alpha'] = [
obj[i]['kitti']['alpha'] for i in range(num_boxes)
]
data_label[img_name]['dims'] = [
obj[i]['kitti']['dimensions'] for i in range(num_boxes)
]
data_label[img_name]['trans'] = [
obj[i]['kitti']['location'] for i in range(num_boxes)
]
data_label[img_name]['rot_y'] = [
obj[i]['kitti']['rotation_y'] for i in range(num_boxes)
]
data_label_list.append(data_label)
with open(os.path.join(DATASET.PKL_FILE), 'wb') as f:
pickle.dump(data_label_list, f, -1)
return data_label_list