def test_single_pbox_set_covar(self):
# Create detection.json file
det_data = {}
det_data['classes'] = self.defaul_det_classes
det_data['img_names'] = self.default_img_names
det_data['detections'] = [[self.default_det_dict]]
det_file_name = os.path.join(self.det_files_root, 'det_file.json')
with open(det_file_name, 'w') as f:
json.dump(det_data, f)
detections = read_files.read_pbox_json(det_file_name, override_cov=200)
self.assertEqual(len(detections), 1)
self.assertIsInstance(detections, read_files.BoxLoader)
img_iterator = iter(detections)
img_dets = next(img_iterator)
self.assertEqual(len(img_dets), 1)
self.assertIsInstance(img_dets[0], PBoxDetInst)
self.assertTrue(
np.allclose(img_dets[0].covs,
[[[200, 0], [0, 200]], [[200, 0], [0, 200]]]))
self.assertTrue(
np.allclose(img_dets[0].box, self.default_det_dict['bbox']))
self.assertListEqual(list(img_dets[0].class_list),
self.default_det_dict['label_probs'])
def load_gt_and_det_data(gt_loc, det_json, data_type):
"""
Function for loading ground-truth and detection instances for a single folder/sequence
:param gt_loc: location of folder/file where ground-truth information can be extracted
:param det_json: detection json file for the given folder/sequence
:param data_type: string defining if we are analysing 'coco' or 'rvc1' data as this will effect
what, and how, data is loaded given the specific filtering process of rvc1.
:return: gt_instances, det_instances, and class list
"""
if data_type == 'coco':
# output is a generator of lists of GroundTruthInstance objects and a map of gt_class_ids
gt_instances, gt_class_ids_map = read_files.read_COCO_gt(gt_loc, ret_classes=True)
# output is a generator of lists of DetectionInstance objects (BBox or PBox depending)
det_instances = read_files.read_pbox_json(det_json, gt_class_ids_map, override_cov=args.set_cov)
class_idxs = [gt_class_ids_map[key] for key in sorted(gt_class_ids_map.keys())]
class_names = list(sorted(gt_class_ids_map.keys()))
class_list = [class_names[idx] for idx in np.argsort(class_idxs)]
elif data_type == 'rvc1':
gt_instances = rvc1_gt_loader.SequenceGTLoader(gt_loc)
det_instances = rvc1_submission_loader.DetSequenceLoader(det_json)
class_list = rvc1_class_list.CLASSES
else:
sys.exit("ERROR! Invalid data type provided")
# check gt_instances and det_instances are the same size
if len(gt_instances) != len(det_instances):
sys.exit("ERROR! gt and det not same length. gt: {0}, det: {1}".format(len(gt_instances), len(det_instances)))
return gt_instances, det_instances, class_list
def test_single_pbox_w_non_gt_class(self):
# Create detection.json file
det_data = {}
det_data['classes'] = self.defaul_det_classes
det_data['img_names'] = self.default_img_names
self.default_det_dict.update(
{'covars': [self.default_covar, self.default_covar]})
self.default_det_dict.update({'label_probs': [0, 0.7, 0.3]})
det_data['detections'] = [[self.default_det_dict]]
det_file_name = os.path.join(self.det_files_root, 'det_file.json')
with open(det_file_name, 'w') as f:
json.dump(det_data, f)
gt_class_ids_dict = {'none': 0, 'diamond': 1, 'square': 2}
detections = read_files.read_pbox_json(det_file_name,
gt_class_ids_dict)
self.assertEqual(len(detections), 1)
self.assertIsInstance(detections, read_files.BoxLoader)
img_iterator = iter(detections)
img_dets = next(img_iterator)
self.assertEqual(len(img_dets), 1)
self.assertIsInstance(img_dets[0], PBoxDetInst)
self.assertTrue(
np.allclose(img_dets[0].covs,
[self.default_covar, self.default_covar]))
self.assertTrue(
np.allclose(img_dets[0].box, self.default_det_dict['bbox']))
expected_label_probs = np.array([0.0, 0.0, 0.7])
self.assertTrue(
np.allclose(img_dets[0].class_list, expected_label_probs))
def test_single_pbox_w_extra_gt_label(self):
# Create detection.json file
det_data = {}
det_data['classes'] = self.defaul_det_classes
det_data['img_names'] = self.default_img_names
self.default_det_dict.update(
{'covars': [self.default_covar, self.default_covar]})
det_data['detections'] = [[self.default_det_dict]]
det_file_name = os.path.join(self.det_files_root, 'det_file.json')
with open(det_file_name, 'w') as f:
json.dump(det_data, f)
detections = read_files.read_pbox_json(det_file_name,
self.default_gt_class_ids_dict)
self.assertEqual(len(detections), 1)
self.assertIsInstance(detections, read_files.BoxLoader)
img_iterator = iter(detections)
img_dets = next(img_iterator)
self.assertEqual(len(img_dets), 1)
self.assertIsInstance(img_dets[0], PBoxDetInst)
self.assertTrue(
np.allclose(img_dets[0].covs,
[self.default_covar, self.default_covar]))
self.assertTrue(
np.allclose(img_dets[0].box, self.default_det_dict['bbox']))
expected_label_probs = [0, 1, 0, 0]
self.assertListEqual(list(img_dets[0].class_list),
expected_label_probs)
def main():
# Load all relevant information for this sequence of information
det_instances, class_list = read_files.read_pbox_json(
args.det_json, override_cov=args.set_cov, get_class_names=True)
img_data_sequence = zip(
sorted(
glob.glob(os.path.join(args.gt_img_folder, '*.' + args.img_type))),
det_instances)
# Go over each image and draw appropriate
for img_name, img_dets in tqdm(img_data_sequence,
total=len(det_instances),
desc='image drawing'):
if args.img_set is None or os.path.basename(img_name) in args.img_set:
save_detection_img(img_name, img_dets, class_list,
args.save_folder, args.corner_mode)
def test_one_pbox_one_bbox_diff_imgs(self):
# Create detection.json file
det_data = {}
det_data['classes'] = self.defaul_det_classes
det_data['img_names'] = self.default_img_names
# Create one bbox detection and one pbox detection
box_det_dict = self.default_det_dict.copy()
self.default_det_dict.update(
{'covars': [self.default_covar, self.default_covar]})
det_data['detections'] = [[self.default_det_dict], [box_det_dict]]
det_file_name = os.path.join(self.det_files_root, 'det_file.json')
with open(det_file_name, 'w') as f:
json.dump(det_data, f)
detections = read_files.read_pbox_json(det_file_name)
self.assertEqual(len(detections), 2)
self.assertIsInstance(detections, read_files.BoxLoader)
# Image 1
img_iterator = iter(detections)
img_dets = next(img_iterator)
self.assertEqual(len(img_dets), 1)
self.assertIsInstance(img_dets[0], PBoxDetInst)
self.assertTrue(
np.allclose(img_dets[0].covs,
[self.default_covar, self.default_covar]))
self.assertTrue(
np.allclose(img_dets[0].box, self.default_det_dict['bbox']))
self.assertListEqual(list(img_dets[0].class_list),
self.default_det_dict['label_probs'])
# Image 2
img_dets = next(img_iterator)
self.assertEqual(len(img_dets), 1)
self.assertIsInstance(img_dets[0], BBoxDetInst)
self.assertTrue(
np.allclose(img_dets[0].box, self.default_det_dict['bbox']))
self.assertListEqual(list(img_dets[0].class_list),
self.default_det_dict['label_probs'])
def gen_param_sequence():
"""
Function for generating the parameter sequence to be used in evaluation procedure.
Parameter sequence holds all GroundTruthInstances, DetectionInstances, and ground-truth filter flags
across all sequences.
:return: param_sequences: ParamSequenceHolder containing all GroundTruthInstances, DetectionInstances,
and ground-truth filter flags across all sequences being evaluated.
len_sequences: list of sequence lengths for all sequences being evaluated.
"""
# Load GTs and Detections as appropriate for different data sets (multiple sequences or one folder)
if args.test_set == 'coco':
# output is a generator of lists of GTInstance objects and a map of gt_class_ids
gt_instances, gt_class_ids = read_files.read_COCO_gt(
coco_gt_file, ret_classes=True, bbox_gt=args.bbox_gt)
det_filename = args.det_loc
# output is a generator of lists of DetectionInstance objects (BBox or PBox depending on detection)
det_instances = read_files.read_pbox_json(det_filename,
gt_class_ids,
override_cov=args.set_cov,
prob_seg=args.prob_seg)
all_gt_instances = [gt_instances]
all_det_instances = [det_instances]
elif args.test_set == 'rvc1':
# output is a list of generator of generators of GTInstance objects
all_gt_instances = rvc1_gt_loader.read_ground_truth(
rvc1_gt_folder, bbox_gt=args.bbox_gt)
# Currently set this version to assume there is a .json file for all sequences in ground-truth
all_det_instances = rvc1_submission_loader.read_submission(
args.det_loc,
["{0:06d}".format(idx) for idx in range(len(all_gt_instances))],
override_cov=args.set_cov)
else:
sys.exit(
"ERROR! Invalid test_set parameter (must be 'coco' or 'rvc1')")
param_sequence = ParamSequenceHolder(all_gt_instances, all_det_instances)
len_sequences = [
len(all_gt_instances[idx]) for idx in range(len(all_gt_instances))
]
return param_sequence, len_sequences
def main():
# Load all relevant information for this sequence of information
det_instances, class_list = read_files.read_pbox_json(
args.det_json, override_cov=args.set_cov, get_class_names=True)
all_images = sorted(
glob.glob(os.path.join(args.gt_img_folder, '*.' + args.img_type)))
if len(all_images) != len(det_instances):
sys.exit(
"ERROR! Ground truth images (--gt_img_folder) and det_instances are not the same length."
"\ngt_img_folder: {0}, det_instances: {1}".format(
len(all_images), len(det_instances)))
img_data_sequence = zip(all_images, det_instances)
# Go over each image and draw appropriate
for img_name, img_dets in tqdm(img_data_sequence,
total=len(det_instances),
desc='image drawing'):
if args.img_set is None or os.path.basename(img_name) in args.img_set:
save_detection_img(img_name, img_dets, class_list,
args.save_folder, args.corner_mode)
def gen_param_sequence():
"""
Function for generating the parameter sequence to be used in evaluation procedure.
Parameter sequence holds all GroundTruthInstances, DetectionInstances, and ground-truth filter flags
across all sequences.
:return: param_sequences: ParamSequenceHolder containing all GroundTruthInstances, DetectionInstances,
and ground-truth filter flags across all sequences being evaluated.
len_sequences: list of sequence lengths for all sequences being evaluated.
"""
# Load GTs and Detections as appropriate for different data sets (multiple sequences or one folder)
#if args.test_set == 'coco':
# output is a generator of lists of GTInstance objects and a map of gt_class_ids
coco_gt_file = 'gt.json'
det_loc = 'pred.json'
gt_instances, gt_class_ids = read_files.read_COCO_gt(coco_gt_file,
ret_classes=True,
bbox_gt=True)
det_filename = det_loc
# output is a generator of lists of DetectionInstance objects (BBox or PBox depending on detection)
det_instances = read_files.read_pbox_json(det_filename,
gt_class_ids,
prob_seg=False)
all_gt_instances = [gt_instances]
all_det_instances = [det_instances]
#print("all_gt = ", all_gt_instances[0], all_det_instances)
param_sequence = ParamSequenceHolder(all_gt_instances, all_det_instances)
len_sequences = [
len(all_gt_instances[idx]) for idx in range(len(all_gt_instances))
]
return param_sequence, len_sequences
