def convert_detections(table_path):
"""Convert detections in `table_path` to metric format.
Args:
table_path: Path to TFRecord file of decoder outputs.
Returns:
(preds, gts): metric_pb2.Objects() of predictions and groundtruths.
"""
img_ids = []
img_id_dict = {}
for serialized in tf.io.tf_record_iterator(table_path):
record = record_pb2.Record()
record.ParseFromString(serialized)
img_id = str(tf.make_ndarray(record.fields["frame_id"]))
img_ids.append(img_id)
np_dict = {k: tf.make_ndarray(v) for k, v in record.fields.items()}
img_id_dict[img_id] = np_dict
preds = metrics_pb2.Objects()
gts = metrics_pb2.Objects()
for img_id in img_ids:
# Extract the underlying context string and timestamp
# from the image id.
#
# TODO(vrv): Consider embedding these values into the decoder output
# individually.
context_name = img_id[2:img_id.rindex("_")]
timestamp = int(img_id[img_id.rindex("_") + 1:-1])
np_dict = img_id_dict[img_id]
pred_bboxes = np_dict["bboxes"] # [max boxes, 7]
pred_scores = np_dict["scores"] # [max_boxes]
gt_bboxes = np_dict["gt_bboxes"] # [num_gt_boxes, 7]
gt_labels = np_dict["gt_labels"] # [num_gt_boxes]
class_ids = np_dict["class_ids"] # [max_boxes]
def _add_box(label, box_vec):
label.box.center_x = box_vec[0]
label.box.center_y = box_vec[1]
label.box.center_z = box_vec[2]
label.box.length = box_vec[3]
label.box.width = box_vec[4]
label.box.height = box_vec[5]
label.box.heading = box_vec[6]
num_gts = gt_bboxes.shape[0]
for gt_idx in range(num_gts):
gt_object = metrics_pb2.Object()
gt_object.context_name = context_name
gt_object.frame_timestamp_micros = timestamp
label = gt_object.object
_add_box(label, gt_bboxes[gt_idx])
label.type = gt_labels[gt_idx]
# We should fill in the difficulty level once we want to measure the
# breakdown by LEVEL.
label.detection_difficulty_level = 0
gts.objects.append(gt_object)
num_pds = pred_bboxes.shape[0]
for pd_idx in range(num_pds):
score = pred_scores[pd_idx]
if score < FLAGS.score_threshold:
continue
pd_object = metrics_pb2.Object()
pd_object.context_name = context_name
pd_object.frame_timestamp_micros = timestamp
pd_object.score = score
label = pd_object.object
_add_box(label, pred_bboxes[pd_idx])
label.type = class_ids[pd_idx]
preds.objects.append(pd_object)
return preds, gts
def main(argv):
if len(argv) > 1:
raise tf.app.UsageError("Too many command-line arguments.")
if FLAGS.decoder_path:
assert not FLAGS.car_decoder_path and not FLAGS.ped_decoder_path \
and not FLAGS.cyc_decoder_path, ("Either provide decoder_path or "
"individual decoders but not both.")
else:
assert FLAGS.car_decoder_path and FLAGS.ped_decoder_path and \
FLAGS.cyc_decoder_path, ("No decoder_path specified. Please supply all "
"individual decoder_paths for labels.")
is_single_decoder_file = FLAGS.decoder_path is not None
if is_single_decoder_file:
list_of_decoder_paths = [FLAGS.decoder_path]
else:
# Note the correspondence between _INCLUDED_KITTI_CLASS_NAMES ordering and
# this list.
list_of_decoder_paths = [
FLAGS.car_decoder_path, FLAGS.ped_decoder_path,
FLAGS.cyc_decoder_path
]
# A list of dictionaries mapping img ids to a dictionary of numpy tensors.
table_data = []
img_ids = []
for table_path in list_of_decoder_paths:
img_id_dict = {}
for serialized in tf.io.tf_record_iterator(table_path):
record = record_pb2.Record()
record.ParseFromString(serialized)
img_id = str(tf.make_ndarray(record.fields["img_id"]))
img_ids.append(img_id)
np_dict = {k: tf.make_ndarray(v) for k, v in record.fields.items()}
img_id_dict[img_id] = np_dict
table_data.append(img_id_dict)
img_ids = list(set(img_ids))
if not tf.io.gfile.exists(FLAGS.output_dir):
tf.io.gfile.mkdir(FLAGS.output_dir)
all_kitti_class_names = kitti_metadata.KITTIMetadata().ClassNames()
calib_data = LoadCalibData(tf.io.gfile.GFile(FLAGS.calib_file, "rb"))
count = 0
for img_id in img_ids:
# Ignore padded samples where the img_ids are empty.
if not img_id:
continue
for table_index, img_id_dict in enumerate(table_data):
if img_id in img_id_dict:
np_dict = img_id_dict[img_id]
(location_cam, dimension_cam, rotation_cam, bboxes_2d, scores,
class_ids) = ExtractNpContent(np_dict,
calib_data[img_id + ".txt"])
if is_single_decoder_file:
valid_labels = _INCLUDED_KITTI_CLASS_NAMES
else:
valid_labels = [_INCLUDED_KITTI_CLASS_NAMES[table_index]]
is_first = table_index == 0
for class_name in valid_labels:
class_mask = (
class_ids == all_kitti_class_names.index(class_name))
ExportKITTIDetection(
FLAGS.output_dir, img_id, location_cam[class_mask],
dimension_cam[class_mask], rotation_cam[class_mask],
bboxes_2d[class_mask], scores[class_mask], class_name,
is_first)
count += 1
tf.logging.info("Total example exported: %d", count)