def run_mots_eval(result_dir, sequences, gt_folder, seqmap_filename):
seqmap, max_frames = load_seqmap(seqmap_filename)
print("Loading ground truth...")
gt = load_gt_sequences(gt_folder, sequences)
print("Loading results...")
results = load_sequences(result_dir, sequences)
print("Compute KITTI tracking eval with simplified matching and MOTSA")
print("Evaluate class: Cars")
results_cars = evaluate_class(gt, results, max_frames, 1)
print("Evaluate class: Pedestrians")
results_ped = evaluate_class(gt, results, max_frames, 2)
print("Results for table (no *)")
print("%.1f" % (results_cars.sMOTSA * 100.0),
"%.1f" % (results_ped.sMOTSA * 100.0),
"%.1f" % (results_cars.MOTSA * 100.0),
"%.1f" % (results_ped.MOTSA * 100.0),
"%.1f" % (results_cars.MOTSP * 100.0),
"%.1f" % (results_ped.MOTSP * 100.0),
results_cars.id_switches,
results_ped.id_switches,
sep="\t")
print("Results for table (*)")
print("%.1f" % (results_cars.sMOTSA_all_ids * 100.0),
"%.1f" % (results_ped.sMOTSA_all_ids * 100.0),
"%.1f" % (results_cars.MOTSA_all_ids * 100.0),
"%.1f" % (results_ped.MOTSA_all_ids * 100.0),
"%.1f" % (results_cars.MOTSP * 100.0),
"%.1f" % (results_ped.MOTSP * 100.0),
results_cars.id_switches_all,
results_ped.id_switches_all,
sep="\t")
def visualize_sequences(sequences,
result_folder,
img_folder,
output_folder,
seqmap_filename,
draw_boxes=False,
create_video=True):
seqmap, max_frames = load_seqmap(seqmap_filename)
for seq_id in sequences:
os.makedirs(output_folder + "/" + seq_id, exist_ok=True)
print("Processing sequence", seq_id)
tracks = load_sequences(result_folder, [seq_id])[seq_id]
max_frames_seq = max_frames[seq_id]
visualize_sequence(seq_id, tracks, max_frames_seq, img_folder,
output_folder, draw_boxes, create_video)
print(sequence)
for frame in objects_per_frame.keys():
for object in objects_per_frame.get(frame):
if not object.track_id == 10000:
track_id = (object.track_id % 1000)
det = {'class': object.class_id}
while tracks_gt.get_num_ids() <= track_id:
tracks_gt.add_empty_track()
tracks_gt.add_to_track(frame, track_id, det, object.mask)
if __name__ == '__main__':
config = Config('./configs/config_default')
list_sequences, max_frames = load_seqmap(config.str('mots_seqmap_file'))
for sequence in list_sequences:
tracks_gt = TrackedSequence(max_frames[sequence] + 1)
import_gt_file('./data/mots_gt/' + sequence + '.txt')
raw_detections = import_detections(config, sequence)
segmentations = import_segmentations(config, sequence)
tracks_gt_seg = TrackedSequence(max_frames[sequence] + 1)
while max_frames[sequence] + 1 > len(raw_detections):
raw_detections.append([])
while max_frames[sequence] + 1 > len(segmentations):
segmentations.append([])
dest='compute_poses',
type=bool,
default=True)
parser.add_argument(
'-compute_point_imgs',
help='compute 3 channel point images using camera parameters',
dest='compute_point_imgs',
type=bool,
default=True)
args = parser.parse_args()
config = Config(args.config)
if args.compute_seg:
refinement_net = refinement_net_init(config.str('segmentor_modelpath'))
list_sequences, _ = load_seqmap(config.str('mots_seqmap_file'))
for sequence in list_sequences:
print(sequence)
sequence_dir = config.dir('data_dir') + 'images/' + sequence + '/'
calibration_params = CalibrationParameters(sequence_dir + sequence +
'.txt')
if args.compute_seg:
print('computing segmentations..')
detections = import_detections(config, sequence)
compute_segmentations(
refinement_net, sequence_dir,
config.dir('segmentations_savedir') + sequence + '/',