def main():
cfg.device = torch.device('cuda')
torch.backends.cudnn.benchmark = False
max_per_image = 100
num_classes = 80 if cfg.dataset == 'coco' else 4
dictionary = np.load(cfg.dictionary_file)
colors = COCO_COLORS if cfg.dataset == 'coco' else DETRAC_COLORS
names = COCO_NAMES if cfg.dataset == 'coco' else DETRAC_NAMES
for j in range(len(names)):
col_ = [c * 255 for c in colors[j]]
colors[j] = tuple(col_)
print('Creating model and recover from checkpoint ...')
if 'hourglass' in cfg.arch:
model = exkp(n=5, nstack=2, dims=[256, 256, 384, 384, 384, 512],
modules=[2, 2, 2, 2, 2, 4], num_classes=num_classes)
elif 'resdcn' in cfg.arch:
model = get_pose_resdcn(num_layers=int(cfg.arch.split('_')[-1]), head_conv=64,
num_classes=num_classes, num_codes=cfg.n_codes)
else:
raise NotImplementedError
model = load_demo_model(model, cfg.ckpt_dir)
model = model.to(cfg.device)
model.eval()
# Loading COCO validation images
if 'train' in cfg.data_type:
annotation_file = '{}/annotations/instances_train2017.json'.format(cfg.data_dir)
cfg.data_type = 'train2017'
elif 'test' in cfg.data_type:
annotation_file = '{}/annotations/image_info_test-dev2017.json'.format(cfg.data_dir)
cfg.data_type = 'test2017'
else:
annotation_file = '{}/annotations/instances_val2017.json'.format(cfg.data_dir)
cfg.data_type = 'val2017'
coco = COCO(annotation_file)
# Load all annotations
# cats = coco.loadCats(coco.getCatIds())
# nms = [cat['name'] for cat in cats]
# catIds = coco.getCatIds(catNms=nms)
# imgIds = np.sort(coco.getImgIds()).tolist()
imgIds = coco.getImgIds()
# annIds = coco.getAnnIds(catIds=catIds)
# all_anns = coco.loadAnns(ids=annIds)
for img_id in imgIds:
img = coco.loadImgs(img_id)[0]
image_path = '%s/coco/%s/%s' % (cfg.data_dir, cfg.data_type, img['file_name'])
w_img = int(img['width'])
h_img = int(img['height'])
if w_img < 1 or h_img < 1:
continue
ann_ids = coco.getAnnIds(imgIds=img_id)
gt_anns = coco.loadAnns(ids=ann_ids)
# plot gt mean and std
# image = cv2.imread(image_path)
# # cv2.ellipse(image, center=(int(contour_mean[0]), int(contour_mean[1])),
# # axes=(int(contour_std[0]), int(contour_std[1])),
# # angle=0, startAngle=0, endAngle=360, color=(0, 255, 0),
# # thickness=2)
# cv2.rectangle(image, pt1=(int(contour_mean[0] - contour_std[0] / 2.), int(contour_mean[1] - contour_std[1] / 2.)),
# pt2=(int(contour_mean[0] + contour_std[0] / 2.), int(contour_mean[1] + contour_std[1] / 2.)),
# color=(0, 255, 0), thickness=2)
# cv2.polylines(image, [fixed_contour.astype(np.int32)], True, (0, 0, 255))
# cv2.rectangle(image, pt1=(int(min(fixed_contour[:, 0])), int(min(fixed_contour[:, 1]))),
# pt2=(int(max(fixed_contour[:, 0])), int(max(fixed_contour[:, 1]))),
# color=(255, 0, 0), thickness=2)
# cv2.imshow('GT segments', image)
# if cv2.waitKey() & 0xFF == ord('q'):
# break
image = cv2.imread(image_path, cv2.IMREAD_COLOR)
if image is None:
continue
print('Loading image of id:', img_id)
# plotting the groundtruth
gt_image = image.copy()
gt_blend_mask = np.zeros(shape=gt_image.shape, dtype=np.uint8)
for ann_ in gt_anns:
if ann_['iscrowd'] == 1:
continue
polygons_ = ann_['segmentation']
use_color_key = COLOR_WORLD[random.randint(1, len(COLOR_WORLD)) - 1]
for poly in polygons_:
poly = np.array(poly).reshape((-1, 2))
cv2.polylines(gt_image, [poly.astype(np.int32)], True,
color=switch_tuple(RGB_DICT[use_color_key]),
thickness=2)
cv2.drawContours(gt_blend_mask, [poly.astype(np.int32)], contourIdx=-1,
color=switch_tuple(RGB_DICT[use_color_key]),
thickness=-1)
original_image = image.copy()
height, width = image.shape[0:2]
padding = 127 if 'hourglass' in cfg.arch else 31
imgs = {}
for scale in cfg.test_scales:
new_height = int(height * scale)
new_width = int(width * scale)
if cfg.img_size > 0:
img_height, img_width = cfg.img_size, cfg.img_size
center = np.array([new_width / 2., new_height / 2.], dtype=np.float32)
scaled_size = max(height, width) * 1.0
scaled_size = np.array([scaled_size, scaled_size], dtype=np.float32)
else:
img_height = (new_height | padding) + 1
img_width = (new_width | padding) + 1
center = np.array([new_width // 2, new_height // 2], dtype=np.float32)
scaled_size = np.array([img_width, img_height], dtype=np.float32)
img = cv2.resize(image, (new_width, new_height))
trans_img = get_affine_transform(center, scaled_size, 0, [img_width, img_height])
img = cv2.warpAffine(img, trans_img, (img_width, img_height))
img = img.astype(np.float32) / 255.
img -= np.array(COCO_MEAN if cfg.dataset == 'coco' else DETRAC_MEAN, dtype=np.float32)[None, None, :]
img /= np.array(COCO_STD if cfg.dataset == 'coco' else DETRAC_STD, dtype=np.float32)[None, None, :]
img = img.transpose(2, 0, 1)[None, :, :, :] # from [H, W, C] to [1, C, H, W]
# if cfg.test_flip:
# img = np.concatenate((img, img[:, :, :, ::-1].copy()), axis=0)
imgs[scale] = {'image': torch.from_numpy(img).float(),
'center': np.array(center),
'scale': np.array(scaled_size),
'fmap_h': np.array(img_height // 4),
'fmap_w': np.array(img_width // 4)}
with torch.no_grad():
segmentations = []
predicted_codes = []
start_time = time.time()
print('Start running model ......')
for scale in imgs:
imgs[scale]['image'] = imgs[scale]['image'].to(cfg.device)
hmap, regs, w_h_, _, _, codes, offsets = model(imgs[scale]['image'])[-1]
output = [hmap, regs, w_h_, codes, offsets]
segms = ctsegm_scale_decode(*output,
torch.from_numpy(dictionary.astype(np.float32)).to(cfg.device),
K=cfg.test_topk)
segms = segms.detach().cpu().numpy().reshape(1, -1, segms.shape[2])[0]
top_preds = {}
code_preds = {}
for j in range(cfg.num_vertices):
segms[:, 2 * j:2 * j + 2] = transform_preds(segms[:, 2 * j:2 * j + 2],
imgs[scale]['center'],
imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
segms[:, cfg.num_vertices * 2:cfg.num_vertices * 2 + 2] = transform_preds(
segms[:, cfg.num_vertices * 2:cfg.num_vertices * 2 + 2],
imgs[scale]['center'],
imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
segms[:, cfg.num_vertices * 2 + 2:cfg.num_vertices * 2 + 4] = transform_preds(
segms[:, cfg.num_vertices * 2 + 2:cfg.num_vertices * 2 + 4],
imgs[scale]['center'],
imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
clses = segms[:, -1]
for j in range(num_classes):
inds = (clses == j)
top_preds[j + 1] = segms[inds, :cfg.num_vertices * 2 + 5].astype(np.float32)
top_preds[j + 1][:, :cfg.num_vertices * 2 + 4] /= scale
segmentations.append(top_preds)
predicted_codes.append(code_preds)
segms_and_scores = {j: np.concatenate([d[j] for d in segmentations], axis=0)
for j in range(1, num_classes + 1)} # a Dict label: segments
scores = np.hstack(
[segms_and_scores[j][:, cfg.num_vertices * 2 + 4] for j in range(1, num_classes + 1)])
print('Image processing time {:.4f} sec, preparing output image ......'.format(time.time() - start_time))
if len(scores) > max_per_image:
kth = len(scores) - max_per_image
thresh = np.partition(scores, kth)[kth]
for j in range(1, num_classes + 1):
keep_inds = (segms_and_scores[j][:, cfg.num_vertices * 2 + 4] >= thresh)
segms_and_scores[j] = segms_and_scores[j][keep_inds]
# Use opencv functions to output
output_image = original_image
blend_mask = np.zeros(shape=output_image.shape, dtype=np.uint8)
counter = 1
for lab in segms_and_scores:
# if cfg.dataset == 'coco':
# if names[lab] not in display_cat and cfg.dataset != 'kins':
# continue
for idx in range(len(segms_and_scores[lab])):
res = segms_and_scores[lab][idx]
contour, bbox, score = res[:-5], res[-5:-1], res[-1]
bbox[0] = np.clip(bbox[0], 0, width - 1)
bbox[1] = np.clip(bbox[1], 0, height - 1)
bbox[2] = np.clip(bbox[2], 0, width - 1)
bbox[3] = np.clip(bbox[3], 0, height - 1)
polygon = contour.reshape((-1, 2))
polygon[:, 0] = np.clip(polygon[:, 0], 0, width - 1)
polygon[:, 1] = np.clip(polygon[:, 1], 0, height - 1)
if score > cfg.detect_thres:
# text = names[lab] + ' %.2f' % score
# label_size = cv2.getTextSize(text, cv2.FONT_HERSHEY_COMPLEX, 0.3, 1)
# text_location = [int(bbox[0]) + 2, int(bbox[1]) + 2,
# int(bbox[0]) + 2 + label_size[0][0],
# int(bbox[1]) + 2 + label_size[0][1]]
# cv2.rectangle(output_image, pt1=(int(bbox[0]), int(bbox[1])),
# pt2=(int(bbox[2]), int(bbox[3])),
# color=colors[lab], thickness=2)
# cv2.rectangle(output_image, pt1=(int(bbox[0]), int(bbox[1])),
# pt2=(int(bbox[2]), int(bbox[3])),
# color=nice_colors[names[lab]], thickness=2)
# cv2.putText(output_image, text, org=(int(text_location[0]), int(text_location[3])),
# fontFace=cv2.FONT_HERSHEY_COMPLEX, thickness=1, fontScale=0.3,
# color=nice_colors[names[lab]])
use_color_key = COLOR_WORLD[random.randint(1, len(COLOR_WORLD)) - 1]
cv2.polylines(output_image, [polygon.astype(np.int32)], True,
color=switch_tuple(RGB_DICT[use_color_key]),
thickness=2)
cv2.drawContours(blend_mask, [polygon.astype(np.int32)], contourIdx=-1,
color=switch_tuple(RGB_DICT[use_color_key]),
thickness=-1)
counter += 1
dst_img = cv2.addWeighted(output_image, 0.4, blend_mask, 0.6, 0)
dst_img[blend_mask == 0] = output_image[blend_mask == 0]
gt_dst_img = cv2.addWeighted(gt_image, 0.4, gt_blend_mask, 0.6, 0)
gt_dst_img[gt_blend_mask == 0] = gt_image[gt_blend_mask == 0]
cat_image = np.concatenate([dst_img, gt_dst_img], axis=1)
cv2.imshow('Frames', cat_image)
if cv2.waitKey() & 0xFF == ord('q'):
break
def main():
cfg.device = torch.device('cuda')
torch.backends.cudnn.benchmark = False
max_per_image = 100
if cfg.dataset == 'coco':
num_classes = 80
colors = COCO_COLORS
names = COCO_NAMES
elif cfg.dataset == 'DETRAC':
num_classes = 3
colors = DETRAC_COLORS
names = DETRAC_NAMES
elif cfg.dataset == 'kins':
num_classes = 7
colors = KINS_COLORS
names = KINS_NAMES
else:
print('Please specify correct dataset name.')
raise NotImplementedError
for j in range(len(names)):
col_ = [c * 255 for c in colors[j]]
colors[j] = tuple(col_)
# Set up parameters for outputing video
output_folder = os.path.join(cfg.root_dir, 'demo')
if not os.path.exists(output_folder):
os.mkdir(output_folder)
width = cfg.video_width
height = cfg.video_height
fps = cfg.video_fps # output video configuration
video_out = cv2.VideoWriter(
os.path.join(output_folder, cfg.output_video_file),
cv2.VideoWriter_fourcc('D', 'I', 'V', 'X'), fps, (width, height))
text_out = open(os.path.join(output_folder, cfg.output_text_file), 'w')
dictionary = np.load(cfg.dictionary_file)
print('Creating model and recover from checkpoint ...')
if 'hourglass' in cfg.arch:
model = exkp(n=5,
nstack=2,
dims=[256, 256, 384, 384, 384, 512],
modules=[2, 2, 2, 2, 2, 4],
num_classes=num_classes)
elif 'resdcn' in cfg.arch:
model = get_pose_resdcn(num_layers=int(cfg.arch.split('_')[-1]),
head_conv=64,
num_classes=num_classes,
num_codes=cfg.n_codes)
else:
raise NotImplementedError
model = load_demo_model(model, cfg.ckpt_dir)
model = model.to(cfg.device)
model.eval()
# Loading images
speed_list = []
frame_list = sorted(os.listdir(cfg.img_dir))
n_frames = len(frame_list)
for frame_id in range(n_frames):
frame_name = frame_list[frame_id]
image_path = os.path.join(cfg.img_dir, frame_name)
image = cv2.imread(image_path)
original_image = image.copy()
height, width = image.shape[0:2]
padding = 127 if 'hourglass' in cfg.arch else 31
imgs = {}
for scale in cfg.test_scales:
new_height = int(height * scale)
new_width = int(width * scale)
if cfg.img_size[0] > 0 and cfg.img_size[1] > 0:
img_height, img_width = cfg.img_size[0], cfg.img_size[1]
center = np.array([new_width / 2., new_height / 2.],
dtype=np.float32)
scaled_size = max(height, width) * 1.0
scaled_size = np.array([scaled_size, scaled_size],
dtype=np.float32)
else:
img_height = (new_height | padding) + 1
img_width = (new_width | padding) + 1
center = np.array([new_width // 2, new_height // 2],
dtype=np.float32)
scaled_size = np.array([img_width, img_height],
dtype=np.float32)
img = cv2.resize(image, (new_width, new_height))
trans_img = get_affine_transform(center, scaled_size, 0,
[img_width, img_height])
img = cv2.warpAffine(img, trans_img, (img_width, img_height))
img = img.astype(np.float32) / 255.
img -= np.array(
COCO_MEAN if cfg.dataset == 'coco' else DETRAC_MEAN,
dtype=np.float32)[None, None, :]
img /= np.array(COCO_STD if cfg.dataset == 'coco' else DETRAC_STD,
dtype=np.float32)[None, None, :]
img = img.transpose(
2, 0, 1)[None, :, :, :] # from [H, W, C] to [1, C, H, W]
imgs[scale] = {
'image': torch.from_numpy(img).float(),
'center': np.array(center),
'scale': np.array(scaled_size),
'fmap_h': np.array(img_height // 4),
'fmap_w': np.array(img_width // 4)
}
with torch.no_grad():
segmentations = []
predicted_codes = []
start_time = time.time()
for scale in imgs:
imgs[scale]['image'] = imgs[scale]['image'].to(cfg.device)
hmap, regs, w_h_, offsets, _, _, codes = model(
imgs[scale]['image'])[-1]
output = [hmap, regs, w_h_, codes, offsets]
segms = ctsegm_scale_decode(
*output,
torch.from_numpy(dictionary.astype(np.float32)).to(
cfg.device),
K=cfg.test_topk)
segms = segms.detach().cpu().numpy().reshape(
1, -1, segms.shape[2])[0]
top_preds = {}
code_preds = {}
for j in range(cfg.num_vertices):
segms[:, 2 * j:2 * j + 2] = transform_preds(
segms[:, 2 * j:2 * j + 2], imgs[scale]['center'],
imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
segms[:, cfg.num_vertices * 2:cfg.num_vertices * 2 +
2] = transform_preds(
segms[:,
cfg.num_vertices * 2:cfg.num_vertices * 2 + 2],
imgs[scale]['center'], imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
segms[:, cfg.num_vertices * 2 + 2:cfg.num_vertices * 2 +
4] = transform_preds(
segms[:, cfg.num_vertices * 2 +
2:cfg.num_vertices * 2 + 4],
imgs[scale]['center'], imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
clses = segms[:, -1]
for j in range(num_classes):
inds = (clses == j)
top_preds[j + 1] = segms[inds, :cfg.num_vertices * 2 +
5].astype(np.float32)
top_preds[j + 1][:, :cfg.num_vertices * 2 + 4] /= scale
segmentations.append(top_preds)
predicted_codes.append(code_preds)
segms_and_scores = {
j: np.concatenate([d[j] for d in segmentations], axis=0)
for j in range(1, num_classes + 1)
} # a Dict label: segments
scores = np.hstack([
segms_and_scores[j][:, cfg.num_vertices * 2 + 4]
for j in range(1, num_classes + 1)
])
if len(scores) > max_per_image:
kth = len(scores) - max_per_image
thresh = np.partition(scores, kth)[kth]
for j in range(1, num_classes + 1):
keep_inds = (segms_and_scores[j][:, cfg.num_vertices * 2 +
4] >= thresh)
segms_and_scores[j] = segms_and_scores[j][keep_inds]
# codes_and_scores[j] = codes_and_scores[j][keep_inds]
# Use opencv functions to output a video
output_image = original_image
blend_mask = np.zeros(shape=output_image.shape, dtype=np.uint8)
counter = 1
for lab in segms_and_scores:
if cfg.dataset == 'coco':
if names[lab] not in display_cat and cfg.dataset != 'kins':
continue
for idx in range(len(segms_and_scores[lab])):
res = segms_and_scores[lab][idx]
contour, bbox, score = res[:-5], res[-5:-1], res[-1]
bbox[0] = np.clip(bbox[0], 0, width - 1)
bbox[1] = np.clip(bbox[1], 0, height - 1)
bbox[2] = np.clip(bbox[2], 0, width - 1)
bbox[3] = np.clip(bbox[3], 0, height - 1)
polygon = contour.reshape((-1, 2))
polygon[:, 0] = np.clip(polygon[:, 0], 0, width - 1)
polygon[:, 1] = np.clip(polygon[:, 1], 0, height - 1)
if score > cfg.detect_thres:
text = names[lab] + ' %.2f' % score
label_size = cv2.getTextSize(text,
cv2.FONT_HERSHEY_COMPLEX,
0.3, 1)
text_location = [
int(bbox[0]) + 2,
int(bbox[1]) + 2,
int(bbox[0]) + 2 + label_size[0][0],
int(bbox[1]) + 2 + label_size[0][1]
]
# cv2.rectangle(output_image, pt1=(int(bbox[0]), int(bbox[1])),
# pt2=(int(bbox[2]), int(bbox[3])),
# color=colors[lab], thickness=2)
# cv2.rectangle(output_image, pt1=(int(bbox[0]), int(bbox[1])),
# pt2=(int(bbox[2]), int(bbox[3])),
# color=nice_colors[names[lab]], thickness=2)
# cv2.putText(output_image, text, org=(int(text_location[0]), int(text_location[3])),
# fontFace=cv2.FONT_HERSHEY_COMPLEX, thickness=1, fontScale=0.3,
# color=nice_colors[names[lab]])
cv2.polylines(output_image, [polygon.astype(np.int32)],
True,
color=nice_colors[names[lab]],
thickness=2)
cv2.drawContours(blend_mask,
[polygon.astype(np.int32)],
contourIdx=-1,
color=nice_colors[names[lab]],
thickness=-1)
# add to text file
new_line = '{0},{1},{2:.3f},{3:.3f},{4:.3f},{5:.3f},{6:.4f}\n'.format(
str(frame_id + 1), counter, int(bbox[0]),
int(bbox[1]),
int(bbox[2]) - int(bbox[0]),
int(bbox[3]) - int(bbox[1]), score)
counter += 1
text_out.write(new_line)
dst_img = cv2.addWeighted(output_image, 0.4, blend_mask, 0.6, 0)
dst_img[blend_mask == 0] = output_image[blend_mask == 0]
output_image = dst_img
cv2.imshow('Frames', output_image)
video_out.write(output_image)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
print('Test frame rate:', 1. / np.mean(speed_list))
def main():
cfg.device = torch.device('cuda')
torch.backends.cudnn.benchmark = False
max_per_image = 100
num_classes = 80 if cfg.dataset == 'coco' else 4
colors = COCO_COLORS if cfg.dataset == 'coco' else DETRAC_COLORS
names = COCO_NAMES if cfg.dataset == 'coco' else DETRAC_NAMES
for j in range(len(names)):
col_ = [c * 255 for c in colors[j]]
colors[j] = tuple(col_)
# Set up parameters for outputing video
output_name = 'demo/'
width = cfg.video_width
height = cfg.video_height
fps = cfg.video_fps # output video configuration
video_out = cv2.VideoWriter(cfg.output_video_dir,
cv2.VideoWriter_fourcc('D', 'I', 'V', 'X'), fps, (width, height))
text_out = open(cfg.output_text_dir, 'w')
print('Creating model and recover from checkpoint ...')
if 'hourglass' in cfg.arch:
model = exkp(n=5, nstack=2, dims=[256, 256, 384, 384, 384, 512],
modules=[2, 2, 2, 2, 2, 4], num_classes=num_classes)
else:
raise NotImplementedError
model = load_demo_model(model, cfg.ckpt_dir)
model = model.to(cfg.device)
model.eval()
# Loading images
speed_list = []
frame_list = sorted(os.listdir(cfg.img_dir))
n_frames = len(frame_list)
for frame_id in range(n_frames):
frame_name = frame_list[frame_id]
image_path = os.path.join(cfg.img_dir, frame_name)
image = cv2.imread(image_path)
original_image = image.copy()
height, width = image.shape[0:2]
padding = 127 if 'hourglass' in cfg.arch else 31
imgs = {}
for scale in cfg.test_scales:
new_height = int(height * scale)
new_width = int(width * scale)
if cfg.img_size > 0:
img_height, img_width = cfg.img_size, cfg.img_size
center = np.array([new_width / 2., new_height / 2.], dtype=np.float32)
scaled_size = max(height, width) * 1.0
scaled_size = np.array([scaled_size, scaled_size], dtype=np.float32)
else:
img_height = (new_height | padding) + 1
img_width = (new_width | padding) + 1
center = np.array([new_width // 2, new_height // 2], dtype=np.float32)
scaled_size = np.array([img_width, img_height], dtype=np.float32)
img = cv2.resize(image, (new_width, new_height))
trans_img = get_affine_transform(center, scaled_size, 0, [img_width, img_height])
img = cv2.warpAffine(img, trans_img, (img_width, img_height))
img = img.astype(np.float32) / 255.
img -= np.array(COCO_MEAN if cfg.dataset == 'coco' else DETRAC_MEAN, dtype=np.float32)[None, None, :]
img /= np.array(COCO_STD if cfg.dataset == 'coco' else DETRAC_STD, dtype=np.float32)[None, None, :]
img = img.transpose(2, 0, 1)[None, :, :, :] # from [H, W, C] to [1, C, H, W]
# if cfg.test_flip:
# img = np.concatenate((img, img[:, :, :, ::-1].copy()), axis=0)
imgs[scale] = {'image': torch.from_numpy(img).float(),
'center': np.array(center),
'scale': np.array(scaled_size),
'fmap_h': np.array(img_height // 4),
'fmap_w': np.array(img_width // 4)}
with torch.no_grad():
detections = []
start_time = time.time()
for scale in imgs:
imgs[scale]['image'] = imgs[scale]['image'].to(cfg.device)
output = model(imgs[scale]['image'])[-1]
dets = ctdet_decode(*output, K=cfg.test_topk)
dets = dets.detach().cpu().numpy().reshape(1, -1, dets.shape[2])[0]
top_preds = {}
dets[:, :2] = transform_preds(dets[:, 0:2],
imgs[scale]['center'],
imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
dets[:, 2:4] = transform_preds(dets[:, 2:4],
imgs[scale]['center'],
imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
cls = dets[:, -1]
for j in range(num_classes):
inds = (cls == j)
top_preds[j + 1] = dets[inds, :5].astype(np.float32)
top_preds[j + 1][:, :4] /= scale
detections.append(top_preds)
bbox_and_scores = {}
for j in range(1, num_classes + 1):
bbox_and_scores[j] = np.concatenate([d[j] for d in detections], axis=0)
if len(cfg.test_scales) > 1:
soft_nms(bbox_and_scores[j], Nt=0.5, method=2)
scores = np.hstack([bbox_and_scores[j][:, 4] for j in range(1, num_classes + 1)])
if len(scores) > max_per_image:
kth = len(scores) - max_per_image
thresh = np.partition(scores, kth)[kth]
for j in range(1, num_classes + 1):
keep_inds = (bbox_and_scores[j][:, 4] >= thresh)
bbox_and_scores[j] = bbox_and_scores[j][keep_inds]
# Use opencv functions to output a video
# output_image = cv2.cvtColor(original_image, cv2.COLOR_BGR2RGB)
speed_list.append(time.time() - start_time)
output_image = original_image
counter = 1
for lab in bbox_and_scores:
if cfg.dataset == 'coco':
if names[lab] not in DETRAC_compatible_names:
continue
for boxes in bbox_and_scores[lab]:
x1, y1, x2, y2, score = boxes
if score > cfg.detect_thres:
text = names[lab] + '%.2f' % score
label_size = cv2.getTextSize(text, cv2.FONT_HERSHEY_COMPLEX, 0.3, 1)
text_location = [x1 + 2, y1 + 2,
x1 + 2 + label_size[0][0],
y1 + 2 + label_size[0][1]]
# cv2.rectangle(output_image, pt1=(int(x1), int(y1)),
# pt2=(int(x2), int(y2)),
# color=colors[lab], thickness=2)
cv2.rectangle(output_image, pt1=(int(x1), int(y1)),
pt2=(int(x2), int(y2)),
color=(0, 255, 0), thickness=2)
# cv2.putText(output_image, text, org=(int(text_location[0]), int(text_location[3])),
# fontFace=cv2.FONT_HERSHEY_COMPLEX, thickness=1, fontScale=0.3,
# color=(0, 0, 255))
# add to text file
new_line = '{0},{1},{2:.3f},{3:.3f},{4:.3f},{5:.3f},{6:.4f}\n'.format(str(frame_id + 1),
counter,
x1, y1, x2 - x1, y2 - y1,
score)
counter += 1
text_out.write(new_line)
cv2.imshow('Frames'.format(frame_id), output_image)
video_out.write(output_image)
if cv2.waitKey(5) & 0xFF == ord('q'):
break
print('Test frame rate:', 1. / np.mean(speed_list))
def main():
# Create Test set labels for DETRAC
detrac_root = cfg.label_dir
dataType = 'Test'
test_images = list()
test_objects = list()
annotation_folder = 'DETRAC-{}-Annotations-XML'.format(dataType)
annotation_path = os.path.join(detrac_root, annotation_folder)
if not os.path.exists(annotation_path):
print('annotation_path not exist')
raise FileNotFoundError
label_file = os.path.join(annotation_path, cfg.video_name + '.xml')
tree = ET.parse(label_file)
root = tree.getroot()
object_list = list()
Box_dict = {}
for obj in root.iter('frame'):
boxes = list()
frame_num = int(obj.attrib['num'])
target_list = obj.find('target_list')
for target in target_list:
bbox = target.find('box').attrib
left = float(bbox['left'])
top = float(bbox['top'])
width = float(bbox['width'])
height = float(bbox['height'])
boxes.append([left, top, left + width,
top + height]) # x1, y1, x2, y2
Box_dict[frame_num] = boxes
cfg.device = torch.device('cuda')
torch.backends.cudnn.benchmark = False
max_per_image = 150
num_classes = 80 if cfg.dataset == 'coco' else 4
colors = COCO_COLORS if cfg.dataset == 'coco' else DETRAC_COLORS
names = COCO_NAMES if cfg.dataset == 'coco' else DETRAC_NAMES
for j in range(len(names)):
col_ = [c * 255 for c in colors[j]]
colors[j] = tuple(col_)
# Set up parameters for outputing video
width = cfg.video_width
height = cfg.video_height
fps = cfg.video_fps # output video configuration
video_out = cv2.VideoWriter(
os.path.join(cfg.root_dir, cfg.video_name + '_compare.mkv'),
cv2.VideoWriter_fourcc('D', 'I', 'V', 'X'), fps, (width, height))
print('Creating model and recover from checkpoint ...')
if 'hourglass' in cfg.arch:
model = exkp(n=5,
nstack=2,
dims=[256, 256, 384, 384, 384, 512],
modules=[2, 2, 2, 2, 2, 4],
num_classes=num_classes)
else:
raise NotImplementedError
model = load_demo_model(model, cfg.ckpt_dir)
model = model.to(cfg.device)
model.eval()
# Loading images
speed_list = []
frame_list = sorted(os.listdir(os.path.join(cfg.img_dir, cfg.video_name)))
n_frames = len(frame_list)
for frame_id in range(n_frames):
frame_n = frame_id + 1
frame_name = frame_list[frame_id]
image_path = os.path.join(cfg.img_dir, cfg.video_name, frame_name)
image = cv2.imread(image_path)
original_image = image.copy()
height, width = image.shape[0:2]
padding = 127 if 'hourglass' in cfg.arch else 31
imgs = {}
for scale in cfg.test_scales:
new_height = int(height * scale)
new_width = int(width * scale)
if cfg.img_size > 0:
img_height, img_width = cfg.img_size, cfg.img_size
center = np.array([new_width / 2., new_height / 2.],
dtype=np.float32)
scaled_size = max(height, width) * 1.0
scaled_size = np.array([scaled_size, scaled_size],
dtype=np.float32)
else:
img_height = (new_height | padding) + 1
img_width = (new_width | padding) + 1
center = np.array([new_width // 2, new_height // 2],
dtype=np.float32)
scaled_size = np.array([img_width, img_height],
dtype=np.float32)
img = cv2.resize(image, (new_width, new_height))
trans_img = get_affine_transform(center, scaled_size, 0,
[img_width, img_height])
img = cv2.warpAffine(img, trans_img, (img_width, img_height))
img = img.astype(np.float32) / 255.
img -= np.array(
COCO_MEAN if cfg.dataset == 'coco' else DETRAC_MEAN,
dtype=np.float32)[None, None, :]
img /= np.array(COCO_STD if cfg.dataset == 'coco' else DETRAC_STD,
dtype=np.float32)[None, None, :]
img = img.transpose(
2, 0, 1)[None, :, :, :] # from [H, W, C] to [1, C, H, W]
# if cfg.test_flip:
# img = np.concatenate((img, img[:, :, :, ::-1].copy()), axis=0)
imgs[scale] = {
'image': torch.from_numpy(img).float(),
'center': np.array(center),
'scale': np.array(scaled_size),
'fmap_h': np.array(img_height // 4),
'fmap_w': np.array(img_width // 4)
}
with torch.no_grad():
detections = []
start_time = time.time()
for scale in imgs:
imgs[scale]['image'] = imgs[scale]['image'].to(cfg.device)
output = model(imgs[scale]['image'])[-1]
dets = ctdet_decode(*output, K=cfg.test_topk)
dets = dets.detach().cpu().numpy().reshape(
1, -1, dets.shape[2])[0]
top_preds = {}
dets[:, :2] = transform_preds(
dets[:, 0:2], imgs[scale]['center'], imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
dets[:, 2:4] = transform_preds(
dets[:, 2:4], imgs[scale]['center'], imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
cls = dets[:, -1]
for j in range(num_classes):
inds = (cls == j)
top_preds[j + 1] = dets[inds, :5].astype(np.float32)
top_preds[j + 1][:, :4] /= scale
detections.append(top_preds)
bbox_and_scores = {}
for j in range(1, num_classes + 1):
bbox_and_scores[j] = np.concatenate([d[j] for d in detections],
axis=0)
if len(cfg.test_scales) > 1:
soft_nms(bbox_and_scores[j], Nt=0.5, method=2)
scores = np.hstack(
[bbox_and_scores[j][:, 4] for j in range(1, num_classes + 1)])
if len(scores) > max_per_image:
kth = len(scores) - max_per_image
thresh = np.partition(scores, kth)[kth]
for j in range(1, num_classes + 1):
keep_inds = (bbox_and_scores[j][:, 4] >= thresh)
bbox_and_scores[j] = bbox_and_scores[j][keep_inds]
# Use opencv functions to output a video
speed_list.append(time.time() - start_time)
output_image = original_image
# Plot the GT boxes
gt_bboxes = Box_dict[frame_n]
for rect in gt_bboxes:
x1, y1, x2, y2 = float(rect[0]), float(rect[1]), float(
rect[2]), float(rect[3])
cv2.rectangle(output_image,
pt1=(int(x1), int(y1)),
pt2=(int(x2), int(y2)),
color=(0, 255, 0),
thickness=2)
counter = 1
for lab in bbox_and_scores:
if cfg.dataset == 'coco':
if names[lab] not in DETRAC_compatible_names:
continue
for boxes in bbox_and_scores[lab]:
x1, y1, x2, y2, score = boxes
if score > cfg.detect_thres:
text = names[lab] + '%.2f' % score
label_size = cv2.getTextSize(text,
cv2.FONT_HERSHEY_COMPLEX,
0.3, 1)
text_location = [
x1 + 2, y1 + 2, x1 + 2 + label_size[0][0],
y1 + 2 + label_size[0][1]
]
cv2.rectangle(output_image,
pt1=(int(x1), int(y1)),
pt2=(int(x2), int(y2)),
color=(0, 0, 255),
thickness=2)
# cv2.putText(output_image, text, org=(int(text_location[0]), int(text_location[3])),
# fontFace=cv2.FONT_HERSHEY_COMPLEX, thickness=1, fontScale=0.3,
# color=(0, 0, 255))
cv2.imshow('Frames'.format(frame_id), output_image)
video_out.write(output_image)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
print('Test frame rate:', 1. / np.mean(speed_list))
def main():
cfg.device = torch.device('cuda')
torch.backends.cudnn.benchmark = False
max_per_image = 100
num_classes = 80 if cfg.dataset == 'coco' else 4
dictionary = np.load(cfg.dictionary_file)
colors = COCO_COLORS if cfg.dataset == 'coco' else DETRAC_COLORS
names = COCO_NAMES if cfg.dataset == 'coco' else DETRAC_NAMES
for j in range(len(names)):
col_ = [c * 255 for c in colors[j]]
colors[j] = tuple(col_)
print('Creating model and recover from checkpoint ...')
if 'hourglass' in cfg.arch:
model = exkp(n=5, nstack=2, dims=[256, 256, 384, 384, 384, 512],
modules=[2, 2, 2, 2, 2, 4], num_classes=num_classes)
elif 'resdcn' in cfg.arch:
model = get_pose_resdcn(num_layers=int(cfg.arch.split('_')[-1]), head_conv=64,
num_classes=num_classes, num_codes=cfg.n_codes)
else:
raise NotImplementedError
model = load_demo_model(model, cfg.ckpt_dir)
model = model.to(cfg.device)
model.eval()
# Loading COCO validation images
if 'train' in cfg.data_type:
annotation_file = '{}/annotations/instances_train2017.json'.format(cfg.data_dir)
cfg.data_type = 'train2017'
else:
annotation_file = '{}/annotations/instances_val2017.json'.format(cfg.data_dir)
cfg.data_type = 'val2017'
coco = COCO(annotation_file)
# Load all annotations
cats = coco.loadCats(coco.getCatIds())
# nms = [cat['name'] for cat in cats]
nms = ['giraffe']
catIds = coco.getCatIds(catNms=nms)
imgIds = coco.getImgIds(catIds=catIds)
annIds = coco.getAnnIds(catIds=catIds)
all_anns = coco.loadAnns(ids=annIds)
for annotation in all_anns:
if annotation['iscrowd'] == 1 or type(annotation['segmentation']) != list or len(
annotation['segmentation']) > 1:
continue
img = coco.loadImgs(annotation['image_id'])[0]
image_path = '%s/images/%s/%s' % (cfg.data_dir, cfg.data_type, img['file_name'])
w_img = int(img['width'])
h_img = int(img['height'])
if w_img < 350 or h_img < 350:
continue
polygons = annotation['segmentation'][0]
gt_bbox = annotation['bbox']
gt_x1, gt_y1, gt_w, gt_h = gt_bbox
contour = np.array(polygons).reshape((-1, 2))
if cv2.contourArea(contour.astype(np.int32)) < 200:
continue
# Downsample the contour to fix number of vertices
fixed_contour = resample(contour, num=cfg.num_vertices)
clockwise_flag = check_clockwise_polygon(fixed_contour)
if not clockwise_flag:
fixed_contour = np.flip(fixed_contour, axis=0)
# else:
# fixed_contour = indexed_shape.copy()
# Indexing from the left-most vertex, argmin x-axis
idx = np.argmin(fixed_contour[:, 0])
indexed_shape = np.concatenate((fixed_contour[idx:, :], fixed_contour[:idx, :]), axis=0)
indexed_shape[:, 0] = np.clip(indexed_shape[:, 0], gt_x1, gt_x1 + gt_w)
indexed_shape[:, 1] = np.clip(indexed_shape[:, 1], gt_y1, gt_y1 + gt_h)
updated_bbox = [np.min(indexed_shape[:, 0]), np.min(indexed_shape[:, 1]),
np.max(indexed_shape[:, 0]), np.max(indexed_shape[:, 1])]
w, h = updated_bbox[2] - updated_bbox[0], updated_bbox[3] - updated_bbox[1]
contour_mean = np.mean(indexed_shape, axis=0)
# contour_std = np.std(indexed_shape, axis=0)
# if contour_std < 1e-6 or contour_std == np.inf or contour_std == np.nan: # invalid shapes
# continue
norm_shape = (indexed_shape - contour_mean) / np.array([w / 2., h / 2.])
gt_codes, _ = fast_ista(norm_shape.reshape((1, -1)), dictionary, lmbda=0.005, max_iter=80)
recon_contour = np.matmul(gt_codes, dictionary).reshape((-1, 2)) * np.array([w / 2., h / 2.])
recon_contour = recon_contour + contour_mean
image = cv2.imread(image_path, cv2.IMREAD_COLOR)
if image is None:
continue
original_image = image.copy()
height, width = image.shape[0:2]
padding = 127 if 'hourglass' in cfg.arch else 31
imgs = {}
for scale in cfg.test_scales:
new_height = int(height * scale)
new_width = int(width * scale)
if cfg.img_size > 0:
img_height, img_width = cfg.img_size, cfg.img_size
center = np.array([new_width / 2., new_height / 2.], dtype=np.float32)
scaled_size = max(height, width) * 1.0
scaled_size = np.array([scaled_size, scaled_size], dtype=np.float32)
else:
img_height = (new_height | padding) + 1
img_width = (new_width | padding) + 1
center = np.array([new_width // 2, new_height // 2], dtype=np.float32)
scaled_size = np.array([img_width, img_height], dtype=np.float32)
img = cv2.resize(image, (new_width, new_height))
trans_img = get_affine_transform(center, scaled_size, 0, [img_width, img_height])
img = cv2.warpAffine(img, trans_img, (img_width, img_height))
img = img.astype(np.float32) / 255.
img -= np.array(COCO_MEAN if cfg.dataset == 'coco' else DETRAC_MEAN, dtype=np.float32)[None, None, :]
img /= np.array(COCO_STD if cfg.dataset == 'coco' else DETRAC_STD, dtype=np.float32)[None, None, :]
img = img.transpose(2, 0, 1)[None, :, :, :] # from [H, W, C] to [1, C, H, W]
# if cfg.test_flip:
# img = np.concatenate((img, img[:, :, :, ::-1].copy()), axis=0)
imgs[scale] = {'image': torch.from_numpy(img).float(),
'center': np.array(center),
'scale': np.array(scaled_size),
'fmap_h': np.array(img_height // 4),
'fmap_w': np.array(img_width // 4)}
with torch.no_grad():
segmentations = []
predicted_codes = []
mass_centers = []
start_time = time.time()
print('Start running model ......')
for scale in imgs:
imgs[scale]['image'] = imgs[scale]['image'].to(cfg.device)
hmap, regs, w_h_, _, _, codes, offsets = model(imgs[scale]['image'])[-1]
output = [hmap, regs, w_h_, codes, offsets]
# segms = ctsegm_scale_decode(*output,
# torch.from_numpy(dictionary.astype(np.float32)).to(cfg.device),
# K=cfg.test_topk)
# print(len(output))
segms, pred_codes, pred_center = ctsegm_scale_decode_debug(*output,
torch.from_numpy(dictionary.astype(np.float32)).to(cfg.device),
K=cfg.test_topk)
segms = segms.detach().cpu().numpy().reshape(1, -1, segms.shape[2])[0]
pred_codes = pred_codes.detach().cpu().numpy().reshape(-1, pred_codes.shape[-1])
pred_center = pred_center.detach().cpu().numpy().reshape(-1, 2)
top_preds = {}
code_preds = {}
center_preds = {}
for j in range(cfg.num_vertices):
segms[:, 2 * j:2 * j + 2] = transform_preds(segms[:, 2 * j:2 * j + 2],
imgs[scale]['center'],
imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
segms[:, cfg.num_vertices * 2:cfg.num_vertices * 2 + 2] = transform_preds(
segms[:, cfg.num_vertices * 2:cfg.num_vertices * 2 + 2],
imgs[scale]['center'],
imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
segms[:, cfg.num_vertices * 2 + 2:cfg.num_vertices * 2 + 4] = transform_preds(
segms[:, cfg.num_vertices * 2 + 2:cfg.num_vertices * 2 + 4],
imgs[scale]['center'],
imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
# For mass center
pred_center = transform_preds(pred_center,
imgs[scale]['center'],
imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
clses = segms[:, -1]
for j in range(num_classes):
inds = (clses == j)
top_preds[j + 1] = segms[inds, :cfg.num_vertices * 2 + 5].astype(np.float32)
top_preds[j + 1][:, :cfg.num_vertices * 2 + 4] /= scale
center_preds[j + 1] = pred_center[inds, :] / scale
code_preds[j + 1] = pred_codes[inds, :]
segmentations.append(top_preds)
predicted_codes.append(code_preds)
mass_centers.append(center_preds)
segms_and_scores = {j: np.concatenate([d[j] for d in segmentations], axis=0)
for j in range(1, num_classes + 1)} # a Dict label: segments
segms_and_codes = {j: np.concatenate([d[j] for d in predicted_codes], axis=0)
for j in range(1, num_classes + 1)}
segms_and_centers = {j: np.concatenate([d[j] for d in mass_centers], axis=0)
for j in range(1, num_classes + 1)}
scores = np.hstack(
[segms_and_scores[j][:, cfg.num_vertices * 2 + 4] for j in range(1, num_classes + 1)])
print('Image processing time {:.4f} sec, preparing output image ......'.format(time.time() - start_time))
if len(scores) > max_per_image:
kth = len(scores) - max_per_image
thresh = np.partition(scores, kth)[kth]
for j in range(1, num_classes + 1):
keep_inds = (segms_and_scores[j][:, cfg.num_vertices * 2 + 4] >= thresh)
segms_and_scores[j] = segms_and_scores[j][keep_inds]
segms_and_codes[j] = segms_and_codes[j][keep_inds]
segms_and_centers[j] = segms_and_centers[j][keep_inds]
# Use opencv functions to output
# output_image = original_image
# blend_mask = np.zeros(shape=output_image.shape, dtype=np.uint8)
counter = 1
for lab in segms_and_scores:
output_image = original_image.copy()
# if cfg.dataset == 'coco':
# if names[lab] not in display_cat and cfg.dataset != 'kins':
# continue
for idx in range(len(segms_and_scores[lab])):
res = segms_and_scores[lab][idx]
p_code = segms_and_codes[lab][idx]
p_center = segms_and_centers[lab][idx]
contour, bbox, score = res[:-5], res[-5:-1], res[-1]
bbox[0] = np.clip(bbox[0], 0, width - 1)
bbox[1] = np.clip(bbox[1], 0, height - 1)
bbox[2] = np.clip(bbox[2], 0, width - 1)
bbox[3] = np.clip(bbox[3], 0, height - 1)
polygon = contour.reshape((-1, 2))
polygon[:, 0] = np.clip(polygon[:, 0], 0, width - 1)
polygon[:, 1] = np.clip(polygon[:, 1], 0, height - 1)
if score > cfg.detect_thres:
# text = names[lab] + ' %.2f' % score
# label_size = cv2.getTextSize(text, cv2.FONT_HERSHEY_COMPLEX, 0.3, 1)
# text_location = [int(bbox[0]) + 2, int(bbox[1]) + 2,
# int(bbox[0]) + 2 + label_size[0][0],
# int(bbox[1]) + 2 + label_size[0][1]]
# cv2.rectangle(output_image, pt1=(int(bbox[0]), int(bbox[1])),
# pt2=(int(bbox[2]), int(bbox[3])),
# color=colors[lab], thickness=2)
# cv2.rectangle(output_image, pt1=(int(bbox[0]), int(bbox[1])),
# pt2=(int(bbox[2]), int(bbox[3])),
# color=nice_colors[names[lab]], thickness=2)
# cv2.putText(output_image, text, org=(int(text_location[0]), int(text_location[3])),
# fontFace=cv2.FONT_HERSHEY_COMPLEX, thickness=1, fontScale=0.3,
# color=nice_colors[names[lab]])
# use_color_key = COLOR_WORLD[random.randint(1, len(COLOR_WORLD)) - 1]
# cv2.polylines(output_image, [polygon.astype(np.int32)], True,
# color=switch_tuple(RGB_DICT[use_color_key]),
# thickness=2)
# cv2.drawContours(blend_mask, [polygon.astype(np.int32)], contourIdx=-1,
# color=switch_tuple(RGB_DICT[use_color_key]),
# thickness=-1)
# plot the polygons/contours
cv2.polylines(output_image, [recon_contour.astype(np.int32)], True,
color=switch_tuple(RGB_DICT['green']), thickness=2)
cv2.polylines(output_image, [polygon.astype(np.int32)], True,
color=switch_tuple(RGB_DICT['red']), thickness=2)
# plot the mass center location
cv2.circle(output_image, tuple(contour_mean.astype(np.int32).tolist()),
radius=9, color=switch_tuple(RGB_DICT['green']), thickness=-1)
cv2.circle(output_image, tuple(p_center.astype(np.int32).tolist()),
radius=9, color=switch_tuple(RGB_DICT['red']), thickness=-1)
# dst_img = cv2.addWeighted(output_image, 0.4, blend_mask, 0.6, 0)
# dst_img[blend_mask == 0] = output_image[blend_mask == 0]
# output_image = dst_img
cv2.imshow('Frames', output_image)
if cv2.waitKey() & 0xFF == ord('q'):
break
counter += 1
# show histogram
fig, (ax1, ax2) = plt.subplots(1, 2)
# plot 1
bins = np.linspace(-2, 2, 30)
ax1.hist(gt_codes.reshape((-1,)).tolist(), bins=bins, color='g', density=False, alpha=0.5)
ax1.hist(p_code.reshape((-1,)).tolist(), bins=bins, color='r', density=False, alpha=0.5)
ax1.legend(['GT Coeffs', 'Pred Coeffs'])
ax1.set_xlabel('Sparse Coefficients')
ax1.set_ylabel('Counts')
ax1.set_title('Histogram of Coefficients')
# plot 2
ax2.plot(gt_codes.reshape((-1,)), 'g*-', linewidth=2, markersize=6)
ax2.plot(p_code.reshape((-1,)), 'ro--', linewidth=1, markersize=5)
ax2.legend(['GT Coeffs', 'Pred Coeffs'])
ax2.set_xlabel('Coefficients Index')
ax2.set_ylabel('Value')
ax2.set_title('Coefficients')
plt.show()
plt.close()
def main():
cfg.device = torch.device('cuda')
torch.backends.cudnn.benchmark = False
max_per_image = 100
if cfg.dataset == 'coco':
num_classes = 80
colors = COCO_COLORS
names = COCO_NAMES
elif cfg.dataset == 'DETRAC':
num_classes = 3
colors = DETRAC_COLORS
names = DETRAC_NAMES
elif cfg.dataset == 'kins':
num_classes = 7
colors = KINS_COLORS
names = KINS_NAMES
else:
print('Please specify correct dataset name.')
raise NotImplementedError
dictionary = np.load(cfg.dictionary_file)
colors = COCO_COLORS if cfg.dataset == 'coco' else DETRAC_COLORS
names = COCO_NAMES if cfg.dataset == 'coco' else DETRAC_NAMES
for j in range(len(names)):
col_ = [c * 255 for c in colors[j]]
colors[j] = tuple(col_)
print('Creating model and recover from checkpoint ...')
if 'hourglass' in cfg.arch:
model = exkp(n=5,
nstack=2,
dims=[256, 256, 384, 384, 384, 512],
modules=[2, 2, 2, 2, 2, 4],
num_classes=num_classes)
elif 'resdcn' in cfg.arch:
model = get_pose_resdcn(num_layers=int(cfg.arch.split('_')[-1]),
head_conv=64,
num_classes=num_classes,
num_codes=cfg.n_codes)
else:
raise NotImplementedError
model = load_demo_model(model, cfg.ckpt_dir)
model = model.to(cfg.device)
model.eval()
# Loading COCO validation images
annotation_file = '{}/annotations/instances_{}.json'.format(
cfg.data_dir, cfg.data_type)
coco = COCO(annotation_file)
# Load all annotations
cats = coco.loadCats(coco.getCatIds())
nms = [cat['name'] for cat in cats]
catIds = coco.getCatIds(catNms=nms)
# imgIds = coco.getImgIds(catIds=catIds)
imgIds = coco.getImgIds()
# annIds = coco.getAnnIds(catIds=catIds)
# all_anns = coco.loadAnns(ids=annIds)
# print(len(imgIds), imgIds)
for id in imgIds:
annt_ids = coco.getAnnIds(imgIds=[id])
annotations_per_img = coco.loadAnns(ids=annt_ids)
# print('All annots: ', len(annotations_per_img), annotations_per_img)
img = coco.loadImgs(id)[0]
image_path = '%s/images/%s/%s' % (cfg.data_dir, cfg.data_type,
img['file_name'])
w_img = int(img['width'])
h_img = int(img['height'])
if w_img < 1 or h_img < 1:
continue
img_original = cv2.imread(image_path)
# img_connect = cv2.imread(image_path)
# img_recon = cv2.imread(image_path)
# print('Image id: ', id)
#
# for annt in annotations_per_img:
# if annt['iscrowd'] == 1 or type(annt['segmentation']) != list:
# continue
#
# polygons = get_connected_polygon_using_mask(annt['segmentation'], (h_img, w_img),
# n_vertices=cfg.num_vertices, closing_max_kernel=60)
# gt_bbox = annt['bbox']
# gt_x1, gt_y1, gt_w, gt_h = gt_bbox
# contour = np.array(polygons).reshape((-1, 2))
#
# # Downsample the contour to fix number of vertices
# if len(contour) > cfg.num_vertices:
# resampled_contour = resample(contour, num=cfg.num_vertices)
# else:
# resampled_contour = turning_angle_resample(contour, cfg.num_vertices)
#
# resampled_contour[:, 0] = np.clip(resampled_contour[:, 0], gt_x1, gt_x1 + gt_w)
# resampled_contour[:, 1] = np.clip(resampled_contour[:, 1], gt_y1, gt_y1 + gt_h)
#
# clockwise_flag = check_clockwise_polygon(resampled_contour)
# if not clockwise_flag:
# fixed_contour = np.flip(resampled_contour, axis=0)
# else:
# fixed_contour = resampled_contour.copy()
#
# # Indexing from the left-most vertex, argmin x-axis
# idx = np.argmin(fixed_contour[:, 0])
# indexed_shape = np.concatenate((fixed_contour[idx:, :], fixed_contour[:idx, :]), axis=0)
#
# x1, y1, x2, y2 = gt_x1, gt_y1, gt_x1 + gt_w, gt_y1 + gt_h
#
# # bbox_width, bbox_height = x2 - x1, y2 - y1
# # bbox = [x1, y1, bbox_width, bbox_height]
# # bbox_center = np.array([(x1 + x2) / 2., (y1 + y2) / 2.])
# bbox_center = np.mean(indexed_shape, axis=0)
#
# centered_shape = indexed_shape - bbox_center
#
# # visualize resampled points with multiple parts in image side by side
# for cnt in range(len(annt['segmentation'])):
# polys = np.array(annt['segmentation'][cnt]).reshape((-1, 2))
# cv2.polylines(img_original, [polys.astype(np.int32)], True, (10, 10, 255), thickness=2)
# # cv2.drawContours(img_original, [polys.astype(np.int32)], contourIdx=-1, color=(10, 10, 255), thickness=-1)
#
# cv2.polylines(img_connect, [indexed_shape.astype(np.int32)], True, (10, 10, 255), thickness=2)
# # cv2.drawContours(img_connect, [indexed_shape.astype(np.int32)], contourIdx=-1, color=(10, 10, 255), thickness=-1)
#
# learned_val_codes, _ = fast_ista(centered_shape.reshape((1, -1)), dictionary,
# lmbda=0.1, max_iter=60)
# recon_contour = np.matmul(learned_val_codes, dictionary).reshape((-1, 2))
# recon_contour = recon_contour + bbox_center
# cv2.polylines(img_recon, [recon_contour.astype(np.int32)], True, (10, 10, 255), thickness=2)
# # cv2.drawContours(img_recon, [recon_contour.astype(np.int32)], contourIdx=-1, color=(10, 10, 255), thickness=-1)
# plot gt mean and std
# image = cv2.imread(image_path)
# # cv2.ellipse(image, center=(int(contour_mean[0]), int(contour_mean[1])),
# # axes=(int(contour_std[0]), int(contour_std[1])),
# # angle=0, startAngle=0, endAngle=360, color=(0, 255, 0),
# # thickness=2)
# cv2.rectangle(image, pt1=(int(contour_mean[0] - contour_std[0] / 2.), int(contour_mean[1] - contour_std[1] / 2.)),
# pt2=(int(contour_mean[0] + contour_std[0] / 2.), int(contour_mean[1] + contour_std[1] / 2.)),
# color=(0, 255, 0), thickness=2)
# cv2.polylines(image, [fixed_contour.astype(np.int32)], True, (0, 0, 255))
# cv2.rectangle(image, pt1=(int(min(fixed_contour[:, 0])), int(min(fixed_contour[:, 1]))),
# pt2=(int(max(fixed_contour[:, 0])), int(max(fixed_contour[:, 1]))),
# color=(255, 0, 0), thickness=2)
# cv2.imshow('GT segments', image)
# if cv2.waitKey() & 0xFF == ord('q'):
# break
image = cv2.imread(image_path)
original_image = image.copy()
height, width = image.shape[0:2]
padding = 127 if 'hourglass' in cfg.arch else 31
imgs = {}
for scale in cfg.test_scales:
new_height = int(height * scale)
new_width = int(width * scale)
if cfg.img_size > 0:
img_height, img_width = cfg.img_size, cfg.img_size
center = np.array([new_width / 2., new_height / 2.],
dtype=np.float32)
scaled_size = max(height, width) * 1.0
scaled_size = np.array([scaled_size, scaled_size],
dtype=np.float32)
else:
img_height = (new_height | padding) + 1
img_width = (new_width | padding) + 1
center = np.array([new_width // 2, new_height // 2],
dtype=np.float32)
scaled_size = np.array([img_width, img_height],
dtype=np.float32)
img = cv2.resize(image, (new_width, new_height))
trans_img = get_affine_transform(center, scaled_size, 0,
[img_width, img_height])
img = cv2.warpAffine(img, trans_img, (img_width, img_height))
img = img.astype(np.float32) / 255.
img -= np.array(
COCO_MEAN if cfg.dataset == 'coco' else DETRAC_MEAN,
dtype=np.float32)[None, None, :]
img /= np.array(COCO_STD if cfg.dataset == 'coco' else DETRAC_STD,
dtype=np.float32)[None, None, :]
img = img.transpose(
2, 0, 1)[None, :, :, :] # from [H, W, C] to [1, C, H, W]
# if cfg.test_flip:
# img = np.concatenate((img, img[:, :, :, ::-1].copy()), axis=0)
imgs[scale] = {
'image': torch.from_numpy(img).float(),
'center': np.array(center),
'scale': np.array(scaled_size),
'fmap_h': np.array(img_height // 4),
'fmap_w': np.array(img_width // 4)
}
with torch.no_grad():
segmentations = []
predicted_codes = []
start_time = time.time()
for scale in imgs:
imgs[scale]['image'] = imgs[scale]['image'].to(cfg.device)
hmap, regs, w_h_, offsets, _, _, codes = model(
imgs[scale]['image'])[-1]
output = [hmap, regs, w_h_, codes, offsets]
segms = ctsegm_scale_decode(
*output,
torch.from_numpy(dictionary.astype(np.float32)).to(
cfg.device),
K=cfg.test_topk)
segms = segms.detach().cpu().numpy().reshape(
1, -1, segms.shape[2])[0]
top_preds = {}
code_preds = {}
for j in range(cfg.num_vertices):
segms[:, 2 * j:2 * j + 2] = transform_preds(
segms[:, 2 * j:2 * j + 2], imgs[scale]['center'],
imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
segms[:, cfg.num_vertices * 2:cfg.num_vertices * 2 +
2] = transform_preds(
segms[:,
cfg.num_vertices * 2:cfg.num_vertices * 2 + 2],
imgs[scale]['center'], imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
segms[:, cfg.num_vertices * 2 + 2:cfg.num_vertices * 2 +
4] = transform_preds(
segms[:, cfg.num_vertices * 2 +
2:cfg.num_vertices * 2 + 4],
imgs[scale]['center'], imgs[scale]['scale'],
(imgs[scale]['fmap_w'], imgs[scale]['fmap_h']))
clses = segms[:, -1]
for j in range(num_classes):
inds = (clses == j)
top_preds[j + 1] = segms[inds, :cfg.num_vertices * 2 +
5].astype(np.float32)
top_preds[j + 1][:, :cfg.num_vertices * 2 + 4] /= scale
segmentations.append(top_preds)
predicted_codes.append(code_preds)
segms_and_scores = {
j: np.concatenate([d[j] for d in segmentations], axis=0)
for j in range(1, num_classes + 1)
} # a Dict label: segments
scores = np.hstack([
segms_and_scores[j][:, cfg.num_vertices * 2 + 4]
for j in range(1, num_classes + 1)
])
if len(scores) > max_per_image:
kth = len(scores) - max_per_image
thresh = np.partition(scores, kth)[kth]
for j in range(1, num_classes + 1):
keep_inds = (segms_and_scores[j][:, cfg.num_vertices * 2 +
4] >= thresh)
segms_and_scores[j] = segms_and_scores[j][keep_inds]
# Use opencv functions to output a video
output_image = original_image
blend_mask = np.zeros(shape=output_image.shape, dtype=np.uint8)
# print(blend_mask.shape)
for lab in segms_and_scores:
for idx in range(len(segms_and_scores[lab])):
res = segms_and_scores[lab][idx]
# c_ = codes_and_scores[lab][idx]
# for res in segms_and_scores[lab]:
contour, bbox, score = res[:-5], res[-5:-1], res[-1]
bbox[0] = np.clip(bbox[0], 0, w_img)
bbox[1] = np.clip(bbox[1], 0, h_img)
bbox[2] = np.clip(bbox[2], 0, w_img)
bbox[3] = np.clip(bbox[3], 0, h_img)
if score > cfg.detect_thres:
text = names[lab] # + ' %.2f' % score
# label_size = cv2.getTextSize(text, cv2.FONT_HERSHEY_COMPLEX, thickness=2, fontScale=0.5)
polygon = contour.reshape((-1, 2))
# print('Shape: Poly -- ', polygon.shape)
# print(polygon)
polygon[:, 0] = np.clip(polygon[:, 0], 0, w_img - 1)
polygon[:, 1] = np.clip(polygon[:, 1], 0, h_img - 1)
# use bb tools to draw predictions
color = random.choice(COLOR_WORLD)
# bb.add(output_image, bbox[0], bbox[1], bbox[2], bbox[3], text, color)
cv2.polylines(output_image, [polygon.astype(np.int32)],
True,
RGB_DICT[color],
thickness=2)
cv2.drawContours(blend_mask,
[polygon.astype(np.int32)],
contourIdx=-1,
color=RGB_DICT[color],
thickness=-1)
# color = (random.randint(0, 255), random.randint(0, 255), random.randint(0, 255))
# contour_mean = np.mean(polygon, axis=0)
# contour_std = np.std(polygon, axis=0)
# center_x, center_y = np.mean(polygon, axis=0).astype(np.int32)
# text_location = [bbox[0] + 1, bbox[1] + 1,
# bbox[1] + label_size[0][0] + 1,
# bbox[0] + label_size[0][1] + 1]
# cv2.rectangle(output_image, pt1=(int(bbox[0]), int(bbox[1])),
# pt2=(int(bbox[2]), int(bbox[3])),
# color=color, thickness=1)
# cv2.rectangle(output_image, pt1=(int(np.min(polygon[:, 0])), int(np.min(polygon[:, 1]))),
# pt2=(int(np.max(polygon[:, 0])), int(np.max(polygon[:, 1]))),
# color=(0, 255, 0), thickness=1)
# cv2.polylines(output_image, [polygon.astype(np.int32)], True, color, thickness=2)
# cv2.putText(output_image, text, org=(int(text_location[0]), int(text_location[3])),
# fontFace=cv2.FONT_HERSHEY_COMPLEX, thickness=2, fontScale=0.5,
# color=(255, 0, 0))
# cv2.putText(output_image, text, org=(int(bbox[0]), int(bbox[1])),
# fontFace=cv2.FONT_HERSHEY_COMPLEX, thickness=1, fontScale=0.5,
# color=color)
# show the histgram for predicted codes
# fig = plt.figure()
# plt.plot(np.arange(cfg.n_codes), c_.reshape((-1,)), color='green',
# marker='o', linestyle='dashed', linewidth=2, markersize=6)
# plt.ylabel('Value of each coefficient')
# plt.xlabel('All predicted {} coefficients'.format(cfg.n_codes))
# plt.title('Distribution of the predicted coefficients for {}'.format(text))
# plt.show()
value = [255, 255, 255]
dst_img = cv2.addWeighted(output_image, 0.5, blend_mask, 0.5, 0)
dst_img[blend_mask == 0] = output_image[blend_mask == 0]
# img_original = cv2.copyMakeBorder(img_original, 0, 0, 0, 10, cv2.BORDER_CONSTANT, None, value)
# img_connect = cv2.copyMakeBorder(img_connect, 0, 0, 10, 10, cv2.BORDER_CONSTANT, None, value)
# img_recon = cv2.copyMakeBorder(img_recon, 0, 0, 10, 10, cv2.BORDER_CONSTANT, None, value)
# dst_img = cv2.copyMakeBorder(dst_img, 0, 0, 10, 0, cv2.BORDER_CONSTANT, None, value)
# im_cat = np.concatenate((img_original, img_connect, img_recon, dst_img), axis=1)
# im_cat = np.concatenate((img_original, img_connect, img_recon), axis=1)
cv2.imshow('segmentation', dst_img)
if cv2.waitKey() & 0xFF == ord('q'):
break
def main():
saver = create_saver(cfg.local_rank, save_dir=cfg.ckpt_dir)
logger = create_logger(cfg.local_rank, save_dir=cfg.log_dir)
summary_writer = create_summary(cfg.local_rank, log_dir=cfg.log_dir)
print = logger.info
print(cfg)
torch.manual_seed(315)
torch.backends.cudnn.benchmark = True # disable this if OOM at beginning of training
num_gpus = torch.cuda.device_count()
if cfg.dist:
cfg.device = torch.device('cuda:%d' % cfg.local_rank)
torch.cuda.set_device(cfg.local_rank)
dist.init_process_group(backend='nccl', init_method='env://',
world_size=num_gpus, rank=cfg.local_rank)
else:
cfg.device = torch.device('cuda:%d' % cfg.device_id)
print('Setting up data...')
# Dataset = COCO if cfg.dataset == 'coco' else PascalVOC
if cfg.dataset.upper() == 'DETRAC':
Dataset = DETRAC
elif cfg.dataset.lower() == 'coco':
Dataset = COCO
else:
raise NotImplementedError
train_dataset = Dataset(cfg.data_dir, 'train', split_ratio=cfg.split_ratio, img_size=cfg.img_size)
train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset,
num_replicas=num_gpus,
rank=cfg.local_rank)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=cfg.batch_size // num_gpus
if cfg.dist else cfg.batch_size,
shuffle=not cfg.dist,
num_workers=cfg.num_workers,
pin_memory=True,
drop_last=True,
sampler=train_sampler if cfg.dist else None)
num_classes = train_dataset.num_classes
if cfg.dataset.upper() == 'DETRAC':
Dataset_eval = DETRAC_eval
elif cfg.dataset.lower() == 'coco':
Dataset_eval = COCO_eval
else:
raise NotImplementedError
val_dataset = Dataset_eval(cfg.data_dir, 'test', test_scales=[1.], test_flip=False)
val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=1,
shuffle=False, num_workers=1, pin_memory=False,
collate_fn=val_dataset.collate_fn)
print('Creating model...')
if 'hourglass' in cfg.arch:
# model = get_hourglass[cfg.arch]
model = exkp(n=5, nstack=2, dims=[256, 256, 384, 384, 384, 512],
modules=[2, 2, 2, 2, 2, 4], num_classes=num_classes)
elif 'resdcn' in cfg.arch:
model = get_pose_net(num_layers=int(cfg.arch.split('_')[-1]), num_classes=train_dataset.num_classes)
else:
raise NotImplementedError
if cfg.dist:
# model = nn.SyncBatchNorm.convert_sync_batchnorm(model)
model = model.to(cfg.device)
model = nn.parallel.DistributedDataParallel(model,
device_ids=[cfg.local_rank, ],
output_device=cfg.local_rank)
else:
model = nn.DataParallel(model).to(cfg.device)
# if os.path.isfile(cfg.pretrain_dir):
# model = load_model(model, cfg.pretrain_dir)
if cfg.pretrain_checkpoint is not None and os.path.isfile(cfg.pretrain_checkpoint):
print('Load pretrain model from ' + cfg.pretrain_checkpoint)
model = load_model(model, cfg.pretrain_checkpoint)
optimizer = torch.optim.Adam(model.parameters(), cfg.lr)
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, cfg.lr_step, gamma=0.1)
def train(epoch):
print('\n Epoch: %d' % epoch)
model.train()
tic = time.perf_counter()
for batch_idx, batch in enumerate(train_loader):
for k in batch:
if k != 'meta':
# print(k)
# print(batch)
batch[k] = batch[k].to(device=cfg.device, non_blocking=True)
outputs = model(batch['image'])
hmap, regs, w_h_ = zip(*outputs)
regs = [_tranpose_and_gather_feature(r, batch['inds']) for r in regs]
w_h_ = [_tranpose_and_gather_feature(r, batch['inds']) for r in w_h_]
hmap_loss = _neg_loss(hmap, batch['hmap'])
reg_loss = _reg_loss(regs, batch['regs'], batch['ind_masks'])
w_h_loss = _reg_loss(w_h_, batch['w_h_'], batch['ind_masks'])
loss = hmap_loss + 1 * reg_loss + 0.1 * w_h_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
if batch_idx % cfg.log_interval == 0:
duration = time.perf_counter() - tic
tic = time.perf_counter()
print('[%d/%d-%d/%d] ' % (epoch, cfg.num_epochs, batch_idx, len(train_loader)) +
' hmap_loss= %.5f reg_loss= %.5f w_h_loss= %.5f' %
(hmap_loss.item(), reg_loss.item(), w_h_loss.item()) +
' (%d samples/sec)' % (cfg.batch_size * cfg.log_interval / duration))
step = len(train_loader) * epoch + batch_idx
summary_writer.add_scalar('hmap_loss', hmap_loss.item(), step)
summary_writer.add_scalar('reg_loss', reg_loss.item(), step)
summary_writer.add_scalar('w_h_loss', w_h_loss.item(), step)
return
def val_map(epoch):
print('\n Val@Epoch: %d' % epoch)
model.eval()
torch.cuda.empty_cache()
max_per_image = 100
results = {}
speed_list = []
with torch.no_grad():
for inputs in val_loader:
img_id, inputs = inputs[0]
start_image_time = time.time()
detections = []
for scale in inputs:
inputs[scale]['image'] = inputs[scale]['image'].to(cfg.device)
output = model(inputs[scale]['image'])[-1]
dets = ctdet_decode(*output, K=cfg.test_topk)
dets = dets.detach().cpu().numpy().reshape(1, -1, dets.shape[2])[0]
top_preds = {}
dets[:, :2] = transform_preds(dets[:, 0:2],
inputs[scale]['center'],
inputs[scale]['scale'],
(inputs[scale]['fmap_w'], inputs[scale]['fmap_h']))
dets[:, 2:4] = transform_preds(dets[:, 2:4],
inputs[scale]['center'],
inputs[scale]['scale'],
(inputs[scale]['fmap_w'], inputs[scale]['fmap_h']))
clses = dets[:, -1]
for j in range(val_dataset.num_classes):
inds = (clses == j)
top_preds[j + 1] = dets[inds, :5].astype(np.float32)
top_preds[j + 1][:, :4] /= scale
detections.append(top_preds)
end_image_time = time.time()
bbox_and_scores = {j: np.concatenate([d[j] for d in detections], axis=0)
for j in range(1, val_dataset.num_classes + 1)}
scores = np.hstack([bbox_and_scores[j][:, 4] for j in range(1, val_dataset.num_classes + 1)])
if len(scores) > max_per_image:
kth = len(scores) - max_per_image
thresh = np.partition(scores, kth)[kth]
for j in range(1, val_dataset.num_classes + 1):
keep_inds = (bbox_and_scores[j][:, 4] >= thresh)
bbox_and_scores[j] = bbox_and_scores[j][keep_inds]
results[img_id] = bbox_and_scores
speed_list.append(end_image_time - start_image_time)
eval_results = val_dataset.run_eval(results, save_dir=cfg.ckpt_dir)
print(eval_results)
summary_writer.add_scalar('val_mAP/mAP', eval_results[0], epoch)
print('Average speed on val set:{:.2f}'.format(1. / np.mean(speed_list)))
print('Starting training...')
for epoch in range(1, cfg.num_epochs + 1):
start = time.time()
train_sampler.set_epoch(epoch)
train(epoch)
if cfg.val_interval > 0 and epoch % cfg.val_interval == 0:
val_map(epoch)
print(saver.save(model.module.state_dict(), 'checkpoint_{}'.format(epoch)))
lr_scheduler.step(epoch)
epoch_time = (time.time() - start) / 3600. / 24.
print('ETA:{:.2f} Days'.format((cfg.num_epochs - epoch) * epoch_time))
summary_writer.close()