# Compute descriptors for both object images and distractor images
image_path_list = [image_dir+imname+'.JPEG' for imname in imlist]
distractor_path_list = [distractor_dir+imname+'.JPEG' for imname in distractor_list]
obj_descriptors = captioner.compute_descriptors(image_path_list)
dis_descriptors = captioner.compute_descriptors(distractor_path_list)
################################################################################
# Test top-1 precision
correct_num = 0
total_num = 0
for n_im in range(num_im):
print('testing image %d / %d' % (n_im, num_im))
imname = imlist[n_im]
for sentence in query_dict[imname]:
# compute test image (target object) score given the description sentence
obj_score = retriever.score_descriptors(obj_descriptors[n_im:n_im+1, :],
sentence, captioner, vocab_dict)[0]
# compute distractor scores given the description sentence
dis_idx = distractor_ids_per_im[imname]
dis_scores = retriever.score_descriptors(dis_descriptors[dis_idx, :],
sentence, captioner, vocab_dict)
# for a retrieval to be correct, the object image must score higher than
# all distractor images
correct_num += np.all(obj_score > dis_scores)
total_num += 1
print('Top-1 precision on the whole test set: %f' % (correct_num/total_num))
################################################################################
num_proposal = candidate_boxes.shape[0]
for n_imcrop in range(num_imcrop):
imcrop_name = imcrop_names[n_imcrop]
if imcrop_name not in query_dict:
continue
gt_bbox = np.array(imcrop_bbox_dict[imcrop_name])
IoUs = retriever.compute_iou(candidate_boxes, gt_bbox)
for n_sentence in range(len(query_dict[imcrop_name])):
sentence = query_dict[imcrop_name][n_sentence]
# Scores for each candidate region
if use_context:
scores = retriever.score_descriptors_context(
descriptors, sentence, context_features_dict[imname],
captioner, vocab_dict)
else:
scores = retriever.score_descriptors(descriptors, sentence,
captioner, vocab_dict)
# Evaluate the correctness of top K predictions
topK_ids = np.argsort(-scores)[:K]
topK_IoUs = IoUs[topK_ids]
# whether the K-th (ranking from high to low) candidate is correct
topK_is_correct = np.zeros(K, dtype=bool)
topK_is_correct[:len(topK_ids)] = (topK_IoUs >=
correct_IoU_threshold)
# whether at least one of the top K candidates is correct
topK_any_correct = (np.cumsum(topK_is_correct) > 0)
topK_correct_num += topK_any_correct
total_num += 1
# print intermediate results during testing
if (n_im + 1) % 1000 == 0:
num_imcrop = len(imcrop_names)
num_proposal = candidate_boxes.shape[0]
for n_imcrop in range(num_imcrop):
imcrop_name = imcrop_names[n_imcrop]
if imcrop_name not in query_dict:
continue
gt_bbox = np.array(imcrop_bbox_dict[imcrop_name])
IoUs = retriever.compute_iou(candidate_boxes, gt_bbox)
for n_sentence in range(len(query_dict[imcrop_name])):
sentence = query_dict[imcrop_name][n_sentence]
# Scores for each candidate region
if use_context:
scores = retriever.score_descriptors_context(descriptors, sentence,
context_features_dict[imname], captioner, vocab_dict)
else:
scores = retriever.score_descriptors(descriptors, sentence,
captioner, vocab_dict)
# Evaluate the correctness of top K predictions
topK_ids = np.argsort(-scores)[:K]
topK_IoUs = IoUs[topK_ids]
# whether the K-th (ranking from high to low) candidate is correct
topK_is_correct = np.zeros(K, dtype=bool)
topK_is_correct[:len(topK_ids)] = (topK_IoUs >= correct_IoU_threshold)
# whether at least one of the top K candidates is correct
topK_any_correct = (np.cumsum(topK_is_correct) > 0)
topK_correct_num += topK_any_correct
total_num += 1
# print intermediate results during testing
if (n_im+1) % 1000 == 0:
print('Recall on first %d test images' % (n_im+1))
]
obj_descriptors = captioner.compute_descriptors(image_path_list)
dis_descriptors = captioner.compute_descriptors(distractor_path_list)
################################################################################
# Test top-1 precision
correct_num = 0
total_num = 0
for n_im in range(num_im):
print('testing image %d / %d' % (n_im, num_im))
imname = imlist[n_im]
for sentence in query_dict[imname]:
# compute test image (target object) score given the description sentence
obj_score = retriever.score_descriptors(
obj_descriptors[n_im:n_im + 1, :], sentence, captioner,
vocab_dict)[0]
# compute distractor scores given the description sentence
dis_idx = distractor_ids_per_im[imname]
dis_scores = retriever.score_descriptors(dis_descriptors[dis_idx, :],
sentence, captioner,
vocab_dict)
# for a retrieval to be correct, the object image must score higher than
# all distractor images
correct_num += np.all(obj_score > dis_scores)
total_num += 1
print('Top-1 precision on the whole test set: %f' % (correct_num / total_num))
################################################################################