def initialize_pos_neg_dataset(train_videos, opts, transform=None, multidomain=True):
"""
Return list of pos and list of neg dataset for each domain.
Args:
train_videos:
opts:
transform:
multidomain:
Returns:
datasets_pos: (list of SLDataset) List length: if multidomain, #videos (or domain). Else: 1
datasets_neg: (list of SLDataset) List length: if multidomain, #videos (or domain). Else: 1
"""
num_videos = len(train_videos['video_names'])
datasets_pos = []
datasets_neg = []
pos_count = 0
neg_count = 0
mean_subtractor = SubtractMeans()
for vid_idx in tqdm(range(num_videos)):
file_name_set = set()
train_db_pos = {
'img_path': [], # list of string
'bboxes': [], # list of ndarray left top coordinate [left top width height]
'labels': [], # list of ndarray #action elements. One hot vector
'score_labels': [], # list of scalar 0 (negative) or 1 (positive)
'vid_idx': [] # list of int. Each video (or domain) index
}
train_db_neg = {
'img_path': [], # list of string
'bboxes': [], # list of ndarray left top coordinate [left top width height]
'labels': [], # list of ndarray #action elements. One hot vector
'score_labels': [], # list of scalar 0 (negative) or 1 (positive)
'vid_idx': [] # list of int. Each video (or domain) index
}
print("Generating dataset from video {}/{} from bench {} (current total (pos-neg): {}-{})...".format(
vid_idx + 1, num_videos, train_videos['bench_names'][vid_idx],
len(train_db_pos['labels']), len(train_db_neg['labels'])))
# print("generating dataset from video " + str(vid_idx + 1) + "/" + str(num_videos) +
# "(current total data (pos-neg): " + str(len(train_db_pos['labels'])) +
# "-" + str(len(train_db_neg['labels'])) + ")")
bench_name = train_videos['bench_names'][vid_idx]
video_name = train_videos['video_names'][vid_idx]
video_path = train_videos['video_paths'][vid_idx]
vid_info = get_video_infos(bench_name, video_path, video_name)
train_db_pos_, train_db_neg_ = get_train_dbs(vid_info, opts)
# separate for each bboxes sample
for sample_idx in range(len(train_db_pos_)):
# for img_path_idx in range(len(train_db_pos_[sample_idx]['score_labels'])):
train_db_pos['img_path'].extend(train_db_pos_[sample_idx]['img_path'])
train_db_pos['bboxes'].extend(train_db_pos_[sample_idx]['bboxes'])
train_db_pos['labels'].extend(train_db_pos_[sample_idx]['labels'])
train_db_pos['score_labels'].extend(train_db_pos_[sample_idx]['score_labels'])
train_db_pos['vid_idx'].extend(np.repeat(vid_idx, len(train_db_pos_[sample_idx]['img_path'])))
pos_count += len(train_db_pos['labels'])
print("Finished generating positive dataset (current total data: {})".format(pos_count))
for sample_idx in range(len(train_db_neg_)):
# for img_path_idx in range(len(train_db_neg_[sample_idx]['score_labels'])):
train_db_neg['img_path'].extend(train_db_neg_[sample_idx]['img_path'])
train_db_neg['bboxes'].extend(train_db_neg_[sample_idx]['bboxes'])
train_db_neg['labels'].extend(train_db_neg_[sample_idx]['labels'])
train_db_neg['score_labels'].extend(train_db_neg_[sample_idx]['score_labels'])
train_db_neg['vid_idx'].extend(np.repeat(vid_idx, len(train_db_neg_[sample_idx]['img_path'])))
neg_count += len(train_db_neg['labels'])
file_name_set.update(train_db_neg['img_path'])
file_name_set.update(train_db_pos['img_path'])
img_path_np_dict = {}
print('Loading images into memory...')
for image_name in tqdm(file_name_set):
im = cv2.imread(image_name)
im, _, _, _ = mean_subtractor.__call__(im)
img_path_np_dict[image_name] = im
print("Finished generating negative dataset (current total data: {})".format(neg_count))
dataset_pos = SLDataset(train_db_pos, transform=transform)
dataset_pos.img_path_np_dict = img_path_np_dict
dataset_neg = SLDataset(train_db_neg, transform=transform)
dataset_neg.img_path_np_dict = img_path_np_dict
if multidomain:
datasets_pos.append(dataset_pos)
datasets_neg.append(dataset_neg)
else:
datasets_pos.extend(dataset_pos)
datasets_neg.extend(dataset_neg)
return datasets_pos, datasets_neg
for c in range(0, 3):
im[y1:y2, x1:x2, c] = (alpha_s * obj_img[:, :, c] +
alpha_l * im[y1:y2, x1:x2, c])
new_im_arr.append(im)
cv2.imwrite(
os.path.join(new_img_save_path, 'color',
file[file.rfind('/') + 1:]), im)
# cv2.imshow('hoy', im)
# key = cv2.waitKey(0) & 0xFF
# if key == ord("q"):
# cv2.destroyAllWindows()
# return
current_x += random.randint(-x_var, x_var)
current_y += random.randint(-y_var, y_var)
current_x = min(current_x, max_x)
current_y = min(current_y, max_y)
current_x = max(current_x, 0)
current_y = max(current_y, 0)
gt_file_path = os.path.join(new_img_save_path, 'groundtruth1.txt')
with open(gt_file_path, "w") as text_file:
for (x, y, w, h) in new_gt_arr:
text_file.write("{},{},{},{}\n".format(x, y, w, h))
return new_im_arr, new_gt_arr
if __name__ == '__main__':
vid_info = get_video_infos('vot15', 'datasets/data/vot15', 'bag')
generate_mot_dataset(vid_info, '../datasets/data/vot15/bagMOT')
def __init__(self, train_videos, opts, transform, args):
self.videos = [] # list of clips dict
self.opts = opts
self.transform = transform
self.args = args
self.RL_steps = self.opts['train']['RL_steps'] # clip length
video_names = train_videos['video_names']
video_paths = train_videos['video_paths']
bench_names = train_videos['bench_names']
vid_idxs = np.random.permutation(len(video_names))
for vid_idx in vid_idxs:
# dict consist of set of clips in ONE video
clips = {
'img_path': [],
'frame_start': [],
'frame_end': [],
'init_bbox': [],
'end_bbox': [],
'vid_idx': [],
}
# Load current training video info
video_name = video_names[vid_idx]
video_path = video_paths[vid_idx]
bench_name = bench_names[vid_idx]
# TODO MAYBE ADD MOT
vid_info = get_video_infos(bench_name, video_path, video_name)
if self.RL_steps is None:
self.RL_steps = len(vid_info['gt']) - 1
vid_clip_starts = [0]
vid_clip_ends = [len(vid_info['gt']) - 1]
else:
vid_clip_starts = np.array(
range(len(vid_info['gt']) - self.RL_steps))
vid_clip_starts = np.random.permutation(vid_clip_starts)
vid_clip_ends = vid_clip_starts + self.RL_steps
# number of clips in one video
num_train_clips = min(opts['train']['rl_num_batches'],
len(vid_clip_starts))
# print("num_train_clips of vid " + str(vid_idx) + ": ", str(num_train_clips))
for clipIdx in range(num_train_clips):
frameStart = vid_clip_starts[clipIdx]
frameEnd = vid_clip_ends[clipIdx]
clips['img_path'].append(
vid_info['img_files'][frameStart:frameEnd])
clips['frame_start'].append(frameStart)
clips['frame_end'].append(frameEnd)
clips['init_bbox'].append(vid_info['gt'][frameStart])
clips['end_bbox'].append(vid_info['gt'][frameEnd])
clips['vid_idx'].append(vid_idx)
if num_train_clips > 0: # small hack
self.videos.append(clips)
self.clip_idx = -1 # hack for reset function
self.vid_idx = 0
self.state = None # current bbox
self.gt = None # end bbox
self.current_img = None # current image frame
self.current_patch = None # current patch (transformed)
self.current_img_idx = 0
self.reset()
def initialize_pos_neg_dataset(train_videos, opts,args, transform=None, multidomain=True):
"""
Return list of pos and list of neg dataset for each domain.
Args:
train_videos:
opts:
transform:
multidomain:
Returns:
datasets_pos: (list of SLDataset) List length: if multidomain, #videos (or domain). Else: 1
datasets_neg: (list of SLDataset) List length: if multidomain, #videos (or domain). Else: 1
"""
# datasets_pos = []
# datasets_neg = []
datasets_pos_neg = []
if train_videos==None:
num_videos=1
else:
num_videos = len(train_videos['video_names'])
t0 = time.time()
for vid_idx in range(num_videos):
train_db = {
'img_path': [], # list of string
'bboxes': [], # list of ndarray left top coordinate [left top width height]
'labels': [], # list of ndarray #action elements. One hot vector
'score_labels': [], # list of scalar 0 (negative) or 1 (positive)
# 'vid_idx': [] # list of int. Each video (or domain) index
}
# train_db_neg = {
# 'img_path': [], # list of string
# 'bboxes': [], # list of ndarray left top coordinate [left top width height]
# 'labels': [], # list of ndarray #action elements. One hot vector
# 'score_labels': [], # list of scalar 0 (negative) or 1 (positive)
# 'vid_idx': [] # list of int. Each video (or domain) index
# }
if train_videos == None:
print("generating dataset from ILSVR dataset...")
# train_db_pos_, train_db_neg_ = get_train_dbs_ILSVR(opts)
if args.train_consecutive:
train_db_pos_neg_ = get_train_dbs_ILSVR_consecutive_frame(opts)
elif args.train_mul_step:
train_db_pos_neg_ = get_train_dbs_mul_step(opts)
else:
train_db_pos_neg_ = get_train_dbs_ILSVR(opts)
else:
# print("generating dataset from video " + str(vid_idx + 1) + "/" + str(num_videos) +
# "(current total data (pos-neg): " + str(len(train_db_pos['labels'])) +
# "-" + str(len(train_db_neg['labels'])) + ")")
print("generating dataset from video " + str(vid_idx + 1) + "/" + str(num_videos) +
"(current total data (pos+neg): " + str(len(train_db['labels'])) + ")")
bench_name = train_videos['bench_names'][vid_idx]
video_name = train_videos['video_names'][vid_idx]
video_path = train_videos['video_paths'][vid_idx]
vid_info = get_video_infos(bench_name, video_path, video_name)
train_db_pos_, train_db_neg_ = get_train_dbs(vid_info, opts)
# separate for each bboxes sample
print("before train_db_pos['img_path'].extend", end=' : ')
print(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
for sample_idx in range(len(train_db_pos_neg_)):
# # for img_path_idx in range(len(train_db_pos_[sample_idx]['score_labels'])):
# train_db['img_path'].append(train_db_pos_neg_[sample_idx]['img_path'])
# train_db['bboxes'].append(train_db_pos_neg_[sample_idx]['bboxes'])
# train_db['labels'].append(train_db_pos_neg_[sample_idx]['labels'])
# train_db['score_labels'].append(train_db_pos_neg_[sample_idx]['score_labels'])
# # train_db['vid_idx'].extend(np.repeat(vid_idx, len(train_db_pos_[sample_idx]['img_path'])))
# # train_db['vid_idx'].append(vid_idx)
train_db['img_path'].extend(train_db_pos_neg_[sample_idx]['img_path'])
train_db['bboxes'].extend(train_db_pos_neg_[sample_idx]['bboxes'])
train_db['labels'].extend(train_db_pos_neg_[sample_idx]['labels'])
train_db['score_labels'].extend(train_db_pos_neg_[sample_idx]['score_labels'])
# if len(train_db_pos_neg_[sample_idx]['bboxes'])!=20:
# print("len(train_db_pos_neg_[sample_idx]['bboxes']): %d, img path: %s"%(
# len(train_db_pos_neg_[sample_idx]['bboxes']),train_db_pos_neg_[sample_idx]['img_path']))
# if len(train_db_pos_neg_[sample_idx]['labels'])!=20:
# print("len(train_db_pos_neg_[sample_idx]['labels']): %d, img path: %s"%(
# len(train_db_pos_neg_[sample_idx]['labels']),train_db_pos_neg_[sample_idx]['img_path']))
# if len(train_db_pos_neg_[sample_idx]['score_labels'])!=20:
# print("len(train_db_pos_neg_[sample_idx]['score_labels']): %d, img path: %s"%(
# len(train_db_pos_neg_[sample_idx]['score_labels']),train_db_pos_neg_[sample_idx]['img_path']))
# print('over debug.')
# print("\nFinish generating positive dataset... (current total data: " + str(len(train_db_pos['labels'])) + ")")
# for sample_idx in range(len(train_db_neg_)):
# # for img_path_idx in range(len(train_db_neg_[sample_idx]['score_labels'])):
# train_db['img_path'].append(train_db_neg_[sample_idx]['img_path'])
# train_db['bboxes'].append(train_db_neg_[sample_idx]['bboxes'])
# train_db['labels'].append(train_db_neg_[sample_idx]['labels'])
# train_db['score_labels'].append(train_db_neg_[sample_idx]['score_labels'])
# # train_db['vid_idx'].extend(np.repeat(vid_idx, len(train_db_neg_[sample_idx]['img_path'])))
# train_db['vid_idx'].append(vid_idx)
# print("\nFinish generating negative dataset... (current total data: " + str(len(train_db_neg['labels'])) + ")")
print("after train_db_neg['img_path'].extend", end=' : ')
print(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
#dataset_pos = SLDataset(train_db_pos, transform=transform)
dataset_pos_neg = SLDataset(train_db, transform=transform)
print("after dataset_pos_neg = SLDataset(train_db", end=' : ')
print(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
# dataset_neg = SLDataset(train_db_neg, transform=transform)
# print("after dataset_neg = SLDataset(train_db_neg", end=' : ')
# print(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
if multidomain:
datasets_pos_neg.append(dataset_pos_neg)
#datasets_neg.append(dataset_neg)
else:
if len(datasets_pos_neg)==0:
datasets_pos_neg.append(dataset_pos_neg)
#datasets_neg.append(dataset_neg)
print("after datasets_pos_neg.append(dataset_pos_neg)", end=' : ')
print(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
else:
# datasets_pos[0].train_db['img_path'].extend(dataset_pos.train_db['img_path'])
# datasets_pos[0].train_db['bboxes'].extend(dataset_pos.train_db['bboxes'])
# datasets_pos[0].train_db['labels'].extend(dataset_pos.train_db['labels'])
# datasets_pos[0].train_db['score_labels'].extend(dataset_pos.train_db['score_labels'])
# datasets_pos[0].train_db['vid_idx'].extend(dataset_pos.train_db['vid_idx'])
#
# datasets_neg[0].train_db['img_path'].extend(dataset_neg.train_db['img_path'])
# datasets_neg[0].train_db['bboxes'].extend(dataset_neg.train_db['bboxes'])
# datasets_neg[0].train_db['labels'].extend(dataset_neg.train_db['labels'])
# datasets_neg[0].train_db['score_labels'].extend(dataset_neg.train_db['score_labels'])
# datasets_neg[0].train_db['vid_idx'].extend(dataset_neg.train_db['vid_idx'])
datasets_pos_neg[0].train_db['img_path'].extend(dataset_pos_neg.train_db['img_path'])
datasets_pos_neg[0].train_db['bboxes'].extend(dataset_pos_neg.train_db['bboxes'])
datasets_pos_neg[0].train_db['labels'].extend(dataset_pos_neg.train_db['labels'])
datasets_pos_neg[0].train_db['score_labels'].extend(dataset_pos_neg.train_db['score_labels'])
# datasets_pos_neg[0].train_db['vid_idx'].extend(dataset_pos_neg.train_db['vid_idx'])
t1 = time.time()
all_time = t1 - t0
all_m = all_time // 60
all_s = all_time % 60
print('time of generating dataset: %d m %d s (%d s)' % (all_m, all_s, all_time))
# return datasets_pos, datasets_neg
return datasets_pos_neg
def initialize_pos_neg_dataset(train_videos,
opts,
transform=None,
multidomain=True):
"""
Return list of pos and list of neg dataset for each domain.
Args:
train_videos:
opts:
transform:
multidomain:
Returns:
datasets_pos: (list of SLDataset) List length: if multidomain, #videos (or domain). Else: 1
datasets_neg: (list of SLDataset) List length: if multidomain, #videos (or domain). Else: 1
"""
num_videos = len(train_videos['video_names'])
datasets_pos = []
datasets_neg = []
for vid_idx in range(num_videos):
train_db_pos = {
'img_path': [], # list of string
'bboxes':
[], # list of ndarray left top coordinate [left top width height]
'labels': [], # list of ndarray #action elements. One hot vector
'score_labels': [], # list of scalar 0 (negative) or 1 (positive)
'vid_idx': [] # list of int. Each video (or domain) index
}
train_db_neg = {
'img_path': [], # list of string
'bboxes':
[], # list of ndarray left top coordinate [left top width height]
'labels': [], # list of ndarray #action elements. One hot vector
'score_labels': [], # list of scalar 0 (negative) or 1 (positive)
'vid_idx': [] # list of int. Each video (or domain) index
}
print("generating dataset from video " + str(vid_idx + 1) + "/" +
str(num_videos) + "(current total data (pos-neg): " +
str(len(train_db_pos['labels'])) + "-" +
str(len(train_db_neg['labels'])) + ")")
bench_name = train_videos['bench_names'][vid_idx]
video_name = train_videos['video_names'][vid_idx]
video_path = train_videos['video_paths'][vid_idx]
vid_info = get_video_infos(bench_name, video_path, video_name)
train_db_pos_, train_db_neg_ = get_train_dbs(vid_info, opts)
# separate for each bboxes sample
for sample_idx in range(len(train_db_pos_)):
# for img_path_idx in range(len(train_db_pos_[sample_idx]['score_labels'])):
train_db_pos['img_path'].extend(
train_db_pos_[sample_idx]['img_path'])
train_db_pos['bboxes'].extend(train_db_pos_[sample_idx]['bboxes'])
train_db_pos['labels'].extend(train_db_pos_[sample_idx]['labels'])
train_db_pos['score_labels'].extend(
train_db_pos_[sample_idx]['score_labels'])
train_db_pos['vid_idx'].extend(
np.repeat(vid_idx, len(train_db_pos_[sample_idx]['img_path'])))
print("Finish generating positive dataset... (current total data: " +
str(len(train_db_pos['labels'])) + ")")
for sample_idx in range(len(train_db_neg_)):
# for img_path_idx in range(len(train_db_neg_[sample_idx]['score_labels'])):
train_db_neg['img_path'].extend(
train_db_neg_[sample_idx]['img_path'])
train_db_neg['bboxes'].extend(train_db_neg_[sample_idx]['bboxes'])
train_db_neg['labels'].extend(train_db_neg_[sample_idx]['labels'])
train_db_neg['score_labels'].extend(
train_db_neg_[sample_idx]['score_labels'])
train_db_neg['vid_idx'].extend(
np.repeat(vid_idx, len(train_db_neg_[sample_idx]['img_path'])))
print("Finish generating negative dataset... (current total data: " +
str(len(train_db_neg['labels'])) + ")")
dataset_pos = SLDataset(train_db_pos, transform=transform)
dataset_neg = SLDataset(train_db_neg, transform=transform)
if multidomain:
datasets_pos.append(dataset_pos)
datasets_neg.append(dataset_neg)
else:
datasets_pos.extend(dataset_pos)
datasets_neg.extend(dataset_neg)
return datasets_pos, datasets_neg