def test_mirror():
image_file = 'P2305'
image = cv.imread('data/rscup/images/{}.png'.format(image_file))
labels = cvtools.readlines(
'data/rscup/labelTxt/{}.txt'.format(image_file))[4:]
bboxes = [
list(map(float,
line.split()[:8])) for line in labels if len(line) > 1
]
draw_img = cvtools.draw_boxes_texts(image.copy(),
bboxes,
box_format='polygon')
cvtools.imwrite(draw_img, 'out/{}.png'.format(image_file))
mirror = cvtools.RandomMirror()
h_image, h_bboxes = horizontal_mirror(image.copy(), np.array(bboxes))
# h_image = np.ascontiguousarray(h_image)
draw_img = cvtools.draw_boxes_texts(h_image,
h_bboxes,
box_format='polygon')
print((h_image == draw_img).all())
cvtools.imwrite(draw_img, 'out/h_{}.png'.format(image_file))
v_image, v_bboxes = vertical_mirror(image.copy(), np.array(bboxes))
draw_img = cvtools.draw_boxes_texts(v_image,
v_bboxes,
box_format='polygon')
cvtools.imwrite(draw_img, 'out/v_{}.png'.format(image_file))
hv_image, hv_bboxes = mirror(image.copy(), bboxes)
draw_img = cvtools.draw_boxes_texts(hv_image,
hv_bboxes,
box_format='polygon')
cvtools.imwrite(draw_img, 'out/hv_{}.png'.format(image_file))
def vis_box(self, save_root, has_box=True, has_segm=True, has_crop=True):
for i, image_name in enumerate(self.imgs):
print('Visualize image %d of %d: %s' %
(i, len(self.imgs), image_name))
# read image
image_file = osp.join(self.img_prefix, image_name)
img = cvtools.imread(image_file)
if img is None:
print('{} is None.'.format(image_file))
continue
imgToObjs = self.img_to_objs[i]
for crop_i, img_coor in enumerate(imgToObjs):
if has_crop:
img = cvtools.draw_boxes_texts(
img, img_coor, colors='random', line_width=2)
objs = imgToObjs[img_coor]
if has_box:
boxes = [obj['bbox'] for obj in objs]
img = cvtools.draw_boxes_texts(
img, boxes, colors=[(255, 0, 0)]*len(boxes), box_format='x1y1wh')
if has_segm:
segms = [obj['segmentation'][0] for obj in objs]
img = cvtools.draw_boxes_texts(
img, segms, box_format='polygen')
to_file = osp.join(save_root, 'images', image_name)
cvtools.imwrite(img, to_file)
def add_coco_bbox(self,
bbox,
cat,
conf=1,
show_txt=True,
img_id='default',
coor_len=4):
bbox = np.array(bbox, dtype=np.int32)
# cat = (int(cat) + 1) % 80
cat = int(cat)
# print('cat', cat, self.names[cat])
c = self.colors[cat][0][0].tolist()
if self.theme == 'white':
c = (255 - np.array(c)).tolist()
txt = '{}{:.1f}'.format(self.names[cat], conf)
font = cv2.FONT_HERSHEY_SIMPLEX
cat_size = cv2.getTextSize(txt, font, 0.5, 2)[0]
if coor_len == 8:
self.imgs[img_id] = cvtools.draw_boxes_texts(self.imgs[img_id],
[bbox], [txt],
box_format='polygon')
elif coor_len == 4:
cv2.rectangle(self.imgs[img_id], (bbox[0], bbox[1]),
(bbox[2], bbox[3]), c, 2)
if show_txt:
cv2.rectangle(self.imgs[img_id],
(bbox[0], bbox[1] - cat_size[1] - 2),
(bbox[0] + cat_size[0], bbox[1] - 2), c, -1)
cv2.putText(self.imgs[img_id],
txt, (bbox[0], bbox[1] - 2),
font,
0.5, (0, 0, 0),
thickness=1,
lineType=cv2.LINE_AA)
def save_crop_labeltxt(self, image_name, img_to_objs, save_root):
crops = []
img_name_no_suffix, img_suffix = osp.splitext(image_name)
for crop_i, img_coor in enumerate(img_to_objs):
crop_objs = img_to_objs[img_coor]
if len(crop_objs) == 0:
continue
# write crop results to txt
txt_name = '_'.join([img_name_no_suffix] + [str(crop_i)] +
list(map(str, img_coor))) + '.txt'
txt_content = ''
for crop_obj in crop_objs:
if len(crop_obj) == 0:
continue
polygen = np.array(crop_obj['segmentation'][0]).reshape(-1, 2)
polygen = polygen - np.array(img_coor[:2]).reshape(-1, 2)
line = list(map(str, polygen.reshape(-1)))
cat = self.COCO.cats[crop_obj['category_id']]['name']
diffcult = str(crop_obj['difficult'])
line.append(cat)
line.append(diffcult)
txt_content += ' '.join(line) + '\n'
cvtools.strwrite(txt_content,
osp.join(save_root, 'labelTxt+crop', txt_name))
if len(crops) > 0:
draw_img = cvtools.draw_boxes_texts(img,
crops,
line_width=3,
box_format='x1y1x2y2')
cvtools.imwrite(
draw_img,
osp.join(save_root, 'images', img_name_no_suffix + '.jpg'))
def test_mirror():
image_file = 'P0641'
image = cv.imread(
osp.join(current_path, 'data/DOTA/images/{}.png'.format(image_file)))
labels = cvtools.readlines(
osp.join(current_path, 'data/DOTA/labelTxt/{}.txt'.format(image_file)))
labels = labels[2:]
bboxes = [
list(map(float,
line.split()[:8])) for line in labels if len(line) > 1
]
draw_img = cvtools.draw_boxes_texts(image.copy(),
bboxes,
box_format='polygon')
cvtools.imwrite(draw_img,
osp.join(current_path, 'out', '{}.png'.format(image_file)))
mirror = cvtools.RandomMirror()
# 测试水平镜像
h_image, h_bboxes = horizontal_mirror(image.copy(), np.array(bboxes))
# h_image = np.ascontiguousarray(h_image)
draw_img = cvtools.draw_boxes_texts(h_image,
h_bboxes,
box_format='polygon')
# print((h_image == draw_img).all())
cvtools.imwrite(
draw_img, osp.join(current_path, 'out', 'h_{}.png'.format(image_file)))
# 测试垂直镜像
v_image, v_bboxes = vertical_mirror(image.copy(), np.array(bboxes))
draw_img = cvtools.draw_boxes_texts(v_image,
v_bboxes,
box_format='polygon')
cvtools.imwrite(
draw_img, osp.join(current_path, 'out', 'v_{}.png'.format(image_file)))
# 测试水平、垂直镜像
hv_image, hv_bboxes = mirror(image.copy(), bboxes)
draw_img = cvtools.draw_boxes_texts(hv_image,
hv_bboxes,
box_format='polygon')
cvtools.imwrite(
draw_img, osp.join(current_path, 'out',
'hv_{}.png'.format(image_file)))
def test_RandomRotate():
image_file = 'P2305'
image = cv.imread('data/rscup/images/{}.png'.format(image_file))
labels = cvtools.readlines('data/rscup/labelTxt/{}.txt'.format(image_file))[4:]
bboxes = [list(map(float, line.split()[:8])) for line in labels if len(line) > 1]
draw_img = cvtools.draw_boxes_texts(image.copy(), bboxes, box_format='polygon')
cvtools.imwrite(draw_img, '{}.png'.format(image_file))
rotate = cvtools.RandomRotate()
rotate_image, rotate_bboxes = rotate(image, bboxes, rotate='rotate_90')
draw_img = cvtools.draw_boxes_texts(rotate_image, rotate_bboxes, box_format='polygon')
cvtools.imwrite(draw_img, '90_{}.png'.format(image_file))
rotate_image, rotate_bboxes = rotate(image, bboxes, rotate='rotate_270')
draw_img = cvtools.draw_boxes_texts(rotate_image, rotate_bboxes, box_format='polygon')
cvtools.imwrite(draw_img, '270_{}.png'.format(image_file))
rotate_image, rotate_bboxes = rotate(image, bboxes, rotate='rotate_180')
draw_img = cvtools.draw_boxes_texts(rotate_image, rotate_bboxes, box_format='polygon')
cvtools.imwrite(draw_img, '180_{}.png'.format(image_file))
def test_RandomRotate():
image_file = 'P0126'
image = cv.imread(
osp.join(current_path, 'data/DOTA/images/{}.png'.format(image_file)))
labels = cvtools.readlines(
osp.join(current_path, 'data/DOTA/labelTxt/{}.txt'.format(image_file)))
labels = labels[2:]
bboxes = [
list(map(float,
line.split()[:8])) for line in labels if len(line) > 1
]
draw_img = cvtools.draw_boxes_texts(image.copy(),
bboxes,
box_format='polygon')
cvtools.imwrite(draw_img,
osp.join(current_path, 'out', '{}.png'.format(image_file)))
rotate = cvtools.RandomRotate()
rotate_image, rotate_bboxes = rotate(image, bboxes, rotate='rotate_90')
draw_img = cvtools.draw_boxes_texts(rotate_image,
rotate_bboxes,
box_format='polygon')
cvtools.imwrite(
draw_img, osp.join(current_path, 'out',
'90_{}.png'.format(image_file)))
rotate_image, rotate_bboxes = rotate(image, bboxes, rotate='rotate_270')
draw_img = cvtools.draw_boxes_texts(rotate_image,
rotate_bboxes,
box_format='polygon')
cvtools.imwrite(
draw_img, osp.join(current_path, 'out',
'270_{}.png'.format(image_file)))
rotate_image, rotate_bboxes = rotate(image, bboxes, rotate='rotate_180')
draw_img = cvtools.draw_boxes_texts(rotate_image,
rotate_bboxes,
box_format='polygon')
cvtools.imwrite(
draw_img, osp.join(current_path, 'out',
'180_{}.png'.format(image_file)))
def vis_objs(self, img, objs, vis='bbox', box_format='x1y1wh'):
for obj in objs:
vis_obj = []
if 'ignore' in obj and obj['ignore'] == 1: continue
if vis in obj: vis_obj = obj[vis]
class_name = self.COCO.cats[obj['category_id']]['name'] \
if 'category_id' in obj else ''
img = cvtools.draw_boxes_texts(img,
vis_obj,
class_name,
box_format=box_format)
return img
def _vis_instances_by_cat(self,
save_root,
vis_cats=None,
vis='bbox',
box_format='x1y1wh'):
catImgs = copy.deepcopy(self.COCO.catToImgs)
catImgs = {cat: set(catImgs[cat]) for cat in catImgs}
for cat_id, image_ids in catImgs.items():
cat_name = self.COCO.cats[cat_id]['name']
if vis_cats is not None and cat_name not in vis_cats:
continue
print('Visualize %s' % cat_name)
if cvtools._DEBUG:
roidb = copy.deepcopy(self.COCO.loadImgs(image_ids))[:10]
else:
roidb = copy.deepcopy(self.COCO.loadImgs(image_ids))
for i, entry in enumerate(roidb):
print('Visualize image %d of %d: %s' %
(i, len(roidb), entry['file_name']))
image_name = entry['file_name']
image_file = osp.join(self.img_prefix, image_name)
img = cv2.imdecode(np.fromfile(image_file, dtype=np.uint8),
cv2.IMREAD_COLOR) # support chinese
# img = cv2.imread(image_file) # not support chinese
image_name = osp.splitext(image_name)[0]
if 'crop' in entry:
img = img[entry['crop'][1]:entry['crop'][3],
entry['crop'][0]:entry['crop'][2]]
image_name = '_'.join([image_name] +
list(map(str, entry['crop'])))
if img is None:
print('{} is None.'.format(image_file))
continue
ann_ids = self.COCO.getAnnIds(imgIds=entry['id'], iscrowd=None)
objs = self.COCO.loadAnns(ann_ids)
for obj in objs:
if obj['category_id'] != cat_id:
continue
if 'ignore' in obj and obj['ignore'] == 1:
continue
vis_obj = []
if vis in obj:
vis_obj = obj[vis]
class_name = [cat_name if 'category_id' in obj else '']
img = cvtools.draw_boxes_texts(img,
vis_obj,
class_name,
box_format=box_format)
# save in jpg format for saving storage
cvtools.imwrite(
img, osp.join(save_root, cat_name, image_name + '.jpg'))
def vis_instances(self,
save_root,
vis='bbox',
box_format='x1y1wh',
by_cat=False):
if by_cat:
self._vis_instances_by_cat(save_root, vis, box_format)
image_ids = self.COCO.getImgIds()
image_ids.sort()
if cvtools._DEBUG:
roidb = copy.deepcopy(self.COCO.loadImgs(image_ids))[:10]
else:
roidb = copy.deepcopy(self.COCO.loadImgs(image_ids))
print('{} images.'.format(len(roidb)))
cvtools.makedirs(save_root)
for i, entry in enumerate(roidb):
print('Visualize image %d of %d: %s' %
(i, len(roidb), entry['file_name']))
image_name = entry['file_name']
image_file = osp.join(self.img_prefix, image_name)
img = cvtools.imread(image_file)
image_name = osp.splitext(image_name)[0]
if 'crop' in entry:
img = img[entry['crop'][1]:entry['crop'][3],
entry['crop'][0]:entry['crop'][2]]
image_name = '_'.join([image_name] +
list(map(str, entry['crop'])))
if img is None:
print('{} is None.'.format(image_file))
continue
ann_ids = self.COCO.getAnnIds(imgIds=entry['id'], iscrowd=None)
objs = self.COCO.loadAnns(ann_ids)
if len(objs) == 0:
continue
# Sanitize bboxes -- some are invalid
for obj in objs:
vis_obj = []
if 'ignore' in obj and obj['ignore'] == 1:
continue
if vis in obj:
vis_obj = obj[vis]
class_name = self.COCO.cats[
obj['category_id']]['name'] if 'category_id' in obj else ''
img = cvtools.draw_boxes_texts(img,
vis_obj,
class_name,
box_format=box_format)
# save in jpg format for saving storage
cvtools.imwrite(img, osp.join(save_root, image_name + '.jpg'))
def vis_dets(self, img_name, result):
dets = result['results']
img = cvtools.imread(img_name)
for cls_id, det_cls in dets.items():
det_cls = det_cls[det_cls[:, -1] > 0.5]
if len(det_cls) == 0: continue
ids = poly_nms.poly_gpu_nms(det_cls, 0.15)
det_cls = det_cls[ids]
if cls_id == 15:
img = cvtools.draw_boxes_texts(img, det_cls[:, :-1],
line_width=2,
box_format="polygon")
img_name = osp.basename(img_name)
to_file = osp.join('/code/CenterNet/exp/ctdet/vis', img_name)
cvtools.imwrite(img, to_file)
def vis(self, img, bbox_result, vis_thr=0.5,
to_file='vis.jpg'):
bboxes = np.vstack(bbox_result)
labels = [
np.full(bbox.shape[0], i, dtype=np.int32)
for i, bbox in enumerate(bbox_result)
]
labels = np.concatenate(labels)
inds = np.where(bboxes[:, -1] > vis_thr)[0]
bboxes = bboxes[inds]
labels = labels[inds]
texts = [self.model.CLASSES[index]+'|'+str(round(bbox[-1], 2))
for index, bbox in zip(labels, bboxes)]
img = cvtools.draw_boxes_texts(
img, bboxes[:, :-1], box_format='polygon', line_width=2)
cvtools.imwrite(img, to_file)
def vis_dets(ret, img_id, dataset, save_by_cat=False, split='val'):
if split != 'test':
img_info = dataset.coco.loadImgs([img_id])[0]
ann_ids = dataset.coco.getAnnIds(imgIds=[img_id])
anns = dataset.coco.loadAnns(ids=ann_ids)
file_name = img_info['file_name']
else:
img_info = dataset.images(img_id)
file_name = img_info['filename']
colors = []
img = read_dota_image(img_info, split)
for cls_id, det_cls in ret.items():
if len(det_cls) == 0: continue
det_cls = det_cls[det_cls[:, -1] > 0.1]
if len(det_cls) == 0: continue
ids = poly_nms.poly_gpu_nms(det_cls.astype(np.float32), 0.15)
det_cls = det_cls[ids]
text = [
dataset.class_name[cls_id] + str(round(score, 2))
for score in det_cls[..., -1]
]
img = cvtools.draw_boxes_texts(img,
det_cls[:, :-1],
texts=None,
line_width=2,
colors=[dataset.voc_color[cls_id - 1]] *
len(det_cls),
box_format="polygon")
crop_str = list(map(str, img_info['crop']))
filename = osp.splitext(file_name)[0]
suffix = osp.splitext(file_name)[1]
save_img_name = osp.join('_'.join([filename] + crop_str) + suffix)
if save_by_cat and split != 'test':
cats = {ann['category_id'] for ann in anns}
for cat in cats:
cat_name = dataset.coco.cats[cat]['name']
file = osp.join(opt.debug_dir, cat_name, save_img_name)
cvtools.imwrite(img, file)
else:
file = osp.join(opt.debug_dir, save_img_name)
cvtools.imwrite(img, file)
def add_bbox_for_paper(self, bbox, cat, img_id='default', coor_len=4):
bbox = np.array(bbox, dtype=np.int32)
# cat = (int(cat) + 1) % 80
cat = int(cat)
# print('cat', cat, self.names[cat])
# c = self.colors[cat][0][0].tolist()
colors = [[249, 47, 47], [53, 253, 8], [7, 233, 252], [235, 8, 249],
[8, 22, 249], [194, 249, 8], [249, 95, 8], [33, 8, 249],
[249, 180, 8], [213, 249, 8], [8, 249, 172], [8, 114, 249],
[8, 114, 249], [8, 114, 249], [8, 114, 249]]
c = colors[cat][::-1]
# if self.theme == 'white':
# c = (255 - np.array(c)).tolist()
if coor_len == 8:
self.imgs[img_id] = cvtools.draw_boxes_texts(self.imgs[img_id],
[bbox],
colors=[c],
line_width=2,
box_format='polygon')
elif coor_len == 4:
cv2.rectangle(self.imgs[img_id], (bbox[0], bbox[1]),
(bbox[2], bbox[3]), c, 2)
def _get_dota_item(self, index):
img_id = self.images[index]
img_info = self.coco.loadImgs(ids=[img_id])[0]
file_name = img_info['file_name']
filename = osp.splitext(file_name)[0]
suffix = osp.splitext(file_name)[1]
crop_str = list(map(str, img_info['crop']))
crop_img_path = osp.join('/code/data/DOTA/crop800_80',
'_'.join([filename] + crop_str) + suffix)
if not osp.isfile(crop_img_path):
img_path = os.path.join('/media/data/DOTA/trainval/images',
file_name)
img = cv2.imread(img_path)
sx, sy, ex, ey = img_info['crop']
img = img[sy:ey + 1, sx:ex + 1]
cv2.imwrite(crop_img_path, img)
else:
img = cv2.imread(crop_img_path)
ann_ids = self.coco.getAnnIds(imgIds=[img_id])
# 如果不deep拷贝，修改了anns就修改了self.coco里的标签
anns = copy.deepcopy(self.coco.loadAnns(ids=ann_ids))
# if True:
if self.opt.debug:
segs = [ann['segmentation'][0] for ann in anns]
cvtools.imwrite(
cvtools.draw_boxes_texts(img.copy(),
segs,
box_format='polygon'),
self.opt.debug_dir + '/{}'.format(file_name))
if self.opt.flip:
try:
hv_flip = cvtools.RandomMirror(both=False)
segs = [ann['segmentation'][0].copy() for ann in anns]
for i, seg in enumerate(segs):
if len(seg) != 8:
segm_hull = cv2.convexHull(np.array(seg).reshape(
-1, 2).astype(np.float32),
clockwise=False)
xywha = cv2.minAreaRect(segm_hull)
segs[i] = cv2.boxPoints(xywha).reshape(-1).tolist()
img, segs = hv_flip(img, segs)
# if True:
if self.opt.debug:
cvtools.imwrite(
cvtools.draw_boxes_texts(img.copy(),
segs,
box_format='polygon'),
self.opt.debug_dir + '/flip_{}'.format(file_name))
for i in range(len(anns)):
anns[i]['segmentation'][0] = list(segs[i])
bbox = cv2.boundingRect(
np.array(segs[i], dtype=np.float32).reshape(-1, 2))
anns[i]['bbox'] = list(bbox)
except Exception as e:
print(e)
return []
if self.opt.rotate:
rotate = cvtools.RandomRotate()
segs = [ann['segmentation'][0].copy() for ann in anns]
for i, seg in enumerate(segs):
if len(seg) != 8:
segm_hull = cv2.convexHull(np.array(seg).reshape(
-1, 2).astype(np.float32),
clockwise=False)
xywha = cv2.minAreaRect(segm_hull)
segs[i] = cv2.boxPoints(xywha).reshape(-1).tolist()
img, segs = rotate(img, segs)
# if True:
if self.opt.debug:
cvtools.imwrite(
cvtools.draw_boxes_texts(img.copy(),
segs,
box_format='polygon'),
self.opt.debug_dir + '/rotate_{}'.format(file_name))
for i in range(len(anns)):
anns[i]['segmentation'][0] = list(segs[i])
bbox = cv2.boundingRect(
np.array(segs[i], dtype=np.float32).reshape(-1, 2))
anns[i]['bbox'] = list(bbox)
height, width = img.shape[0], img.shape[1]
c = np.array([img.shape[1] / 2., img.shape[0] / 2.], dtype=np.float32)
# self.opt.input_h = self.opt.input_w = 32 * random.randint(12, 20)
if self.opt.keep_res:
input_h = (height | self.opt.pad) + 1
input_w = (width | self.opt.pad) + 1
s = np.array([input_w, input_h], dtype=np.float32)
else:
s = max(img.shape[0], img.shape[1]) * 1.0
input_h, input_w = self.opt.input_h, self.opt.input_w
# flipped = False
if 'train' in self.split:
if not self.opt.not_rand_crop:
s = s * np.random.choice(np.arange(0.6, 1.4, 0.1))
w_border = self._get_border(128, img.shape[1])
h_border = self._get_border(128, img.shape[0])
c[0] = np.random.randint(low=w_border,
high=img.shape[1] - w_border)
c[1] = np.random.randint(low=h_border,
high=img.shape[0] - h_border)
else:
sf = self.opt.scale
cf = self.opt.shift
c[0] += s * np.clip(np.random.randn() * cf, -2 * cf, 2 * cf)
c[1] += s * np.clip(np.random.randn() * cf, -2 * cf, 2 * cf)
s = s * np.clip(np.random.randn() * sf + 1, 1 - sf, 1 + sf)
# if np.random.random() < self.opt.flip:
# flipped = True
# img = img[:, ::-1, :]
# c[0] = width - c[0] - 1
trans_input = get_affine_transform(c, s, 0, [input_w, input_h])
inp = cv2.warpAffine(img,
trans_input, (input_w, input_h),
flags=cv2.INTER_LINEAR)
gt_boxes = np.array(
[cvtools.x1y1wh_to_x1y1x2y2(ann['bbox']) for ann in anns])
img_box = cvtools.xywh_to_x1y1x2y2(np.array([[c[0], c[1], s, s]]))
img_box[0, 0::2] = np.clip(img_box[0, 0::2], 0, width - 1)
img_box[0, 1::2] = np.clip(img_box[0, 1::2], 0, height - 1)
iofs = cvtools.bbox_overlaps(gt_boxes, img_box, mode='iof')
ids = np.where(iofs > 0.7)[0]
if len(ids) == 0: return []
anns = [anns[ind] for ind in ids]
# if True:
if self.opt.debug:
segs = [ann['segmentation'][0].copy()
for ann in anns] # 复制一份，否则是原视图
inp_draw = inp.copy()
for k in range(len(segs)):
seg = segs[k]
for i in range(0, len(seg), 2):
seg[i:i + 2] = affine_transform(seg[i:i + 2], trans_input)
# seg[i] = np.clip(seg[i], 0, input_w - 1)
# seg[i + 1] = np.clip(seg[i + 1], 0, input_h - 1)
segm_hull = cv2.convexHull(np.array(seg).reshape(-1, 2).astype(
np.float32),
clockwise=False)
xy, _, _ = cv2.minAreaRect(segm_hull)
cv2.circle(inp_draw, (int(xy[0]), int(xy[1])),
radius=5,
color=(0, 0, 255),
thickness=-1)
cvtools.imwrite(
cvtools.draw_boxes_texts(inp_draw,
segs,
draw_start=False,
box_format='polygon'),
osp.join(self.opt.debug_dir, 'trans_' + file_name))
inp = (inp.astype(np.float32) / 255.)
if 'train' in self.split and not self.opt.no_color_aug:
color_aug(self._data_rng, inp, self._eig_val, self._eig_vec)
inp = (inp - self.mean) / self.std
inp = inp.transpose(2, 0, 1)
output_h = input_h // self.opt.down_ratio
output_w = input_w // self.opt.down_ratio
num_classes = self.num_classes
trans_output = get_affine_transform(c, s, 0, [output_w, output_h])
rets = []
# out_size = []
# for down_ratio in down_ratios:
# output_h = input_h // down_ratio
# output_w = input_w // down_ratio
# num_classes = self.num_classes
# out_size.append([output_w, output_h])
# # trans_output = get_affine_transform(c, s, 0, [output_w, output_h])
#
# hm = np.zeros((num_classes, output_h, output_w), dtype=np.float32)
# wh = np.zeros((self.max_objs, 2), dtype=np.float32)
# if self.opt.a_method == 2:
# angle = np.full((self.max_objs, 1), 0.5, dtype=np.float32)
# else:
# angle = np.zeros((self.max_objs, 1), dtype=np.float32)
# dense_wh = np.zeros((2, output_h, output_w), dtype=np.float32)
# reg = np.zeros((self.max_objs, 2), dtype=np.float32)
# ind = np.zeros((self.max_objs), dtype=np.int64)
# reg_mask = np.zeros((self.max_objs), dtype=np.uint8)
# cat_spec_wh = np.zeros((self.max_objs, num_classes * 2), dtype=np.float32)
# cat_spec_mask = np.zeros((self.max_objs, num_classes * 2), dtype=np.uint8)
#
# # for k in range(num_objs):
# # cls_id = int(self.cat_ids[anns[k]['category_id']])
# # draw_heatmap(hm[cls_id], osp.join(self.opt.debug_dir, 'heatmap_' + str(cls_id) + '_' + file_name))
# ret = {'input': inp, 'hm': hm, 'reg_mask': reg_mask, 'ind': ind, 'wh': wh, 'a': angle}
# if self.opt.dense_wh:
# hm_a = hm.max(axis=0, keepdims=True)
# dense_wh_mask = np.concatenate([hm_a, hm_a], axis=0)
# ret.update({'dense_wh': dense_wh, 'dense_wh_mask': dense_wh_mask})
# del ret['wh']
# elif self.opt.cat_spec_wh:
# ret.update({'cat_spec_wh': cat_spec_wh, 'cat_spec_mask': cat_spec_mask})
# del ret['wh']
# if self.opt.reg_offset:
# ret.update({'reg': reg})
# # if self.opt.debug > 0 or not self.split == 'train':
# # gt_det = np.array(gt_det, dtype=np.float32) if len(gt_det) > 0 else \
# # np.zeros((1, 6), dtype=np.float32)
# # meta = {'c': c, 's': s, 'gt_det': gt_det, 'img_id': img_id}
# # ret['meta'] = meta
# rets.append(ret)
#
# draw_gaussian = draw_msra_gaussian if self.opt.mse_loss else \
# draw_umich_gaussian
# if not self.opt.fpn:
# output_w, output_h = out_size[0]
# trans_output = get_affine_transform(c, s, 0, out_size[0])
#
# for k in range(num_objs):
# ann = anns[k]
# cls_id = int(self.cat_ids[ann['category_id']])
#
# # 确定GT分配给哪个FPN层
# if self.opt.fpn:
# bbox = ann['bbox']
# fpn_k = int(math.log(224. / math.sqrt(bbox[2] * bbox[3]), 2))
# if fpn_k < 0:
# fpn_k = 0
# if fpn_k > 2:
# fpn_k = 2
# ret = rets[fpn_k]
# output_w, output_h = out_size[fpn_k]
# trans_output = get_affine_transform(c, s, 0, out_size[fpn_k])
#
# segm = np.array(ann['segmentation'][0])
# # if flipped:
# # for i in range(0, len(segm), 2):
# # segm[i] = width - segm[i] - 1
# for i in range(0, len(segm), 2):
# segm[i:i + 2] = affine_transform(segm[i:i + 2], trans_output)
# segm[i] = np.clip(segm[i], 0, output_w - 1)
# segm[i + 1] = np.clip(segm[i + 1], 0, output_h - 1)
#
# segm_hull = cv2.convexHull(segm.reshape(-1, 2).astype(np.float32),
# clockwise=False)
# xy, (w, h), a = cv2.minAreaRect(segm_hull)
# hm = ret['hm']
# reg_mask = ret['reg_mask']
# ind = ret['ind']
# wh = ret['wh']
# angle = ret['a']
# if h > 0 and w > 0:
# a, w, h = convert_angle(a, w, h, self.opt.a_method)
# ct = np.array(xy, dtype=np.float32)
# ct_int = ct.astype(np.int32)
# radius = gaussian_radius((math.ceil(h), math.ceil(w)))
# radius = max(0, int(radius))
# radius = self.opt.hm_gauss if self.opt.mse_loss else radius
# # radius = np.array((h / 3., w / 3.), np.int32)
# draw_gaussian(hm[cls_id], ct_int, radius)
# wh[k] = 1. * w, 1. * h
# gt_a = a / 90.
# if self.opt.a_method == 2:
# gt_a = (a + 90.) / 180.
# angle[k] = gt_a
# ind[k] = ct_int[1] * output_w + ct_int[0]
# if 'reg' in ret:
# ret['reg'][k] = ct - ct_int
# reg_mask[k] = 1
# if 'cat_spec_wh' in ret:
# ret['cat_spec_wh'][k, cls_id * 2: cls_id * 2 + 2] = wh[k]
# if 'cat_spec_mask' in ret:
# ret['cat_spec_mask'][k, cls_id * 2: cls_id * 2 + 2] = 1
# if self.opt.dense_wh:
# draw_dense_reg(ret['dense_wh'], hm.max(axis=0), ct_int,
# wh[k], radius)
draw_gaussian = draw_msra_gaussian if self.opt.mse_loss else \
draw_umich_gaussian
hm = np.zeros((num_classes, output_h, output_w), dtype=np.float32)
wh = np.zeros((self.max_objs, 2), dtype=np.float32)
if self.opt.a_method == 2:
angle = np.full((self.max_objs, 1), 0.5, dtype=np.float32)
else:
angle = np.zeros((self.max_objs, 1), dtype=np.float32)
dense_wh = np.zeros((2, output_h, output_w), dtype=np.float32)
reg = np.zeros((self.max_objs, 2), dtype=np.float32)
ind = np.zeros((self.max_objs), dtype=np.int64)
reg_mask = np.zeros((self.max_objs), dtype=np.uint8)
cat_spec_wh = np.zeros((self.max_objs, num_classes * 2),
dtype=np.float32)
cat_spec_mask = np.zeros((self.max_objs, num_classes * 2),
dtype=np.uint8)
anns = re_anns(anns, trans_output, output_w, output_h)
num_objs = min(len(anns), self.max_objs)
# if True:
if self.opt.debug:
gt_img = cv2.warpAffine(img,
trans_output, (output_w, output_h),
flags=cv2.INTER_LINEAR)
segs = [ann['segmentation'][0] for ann in anns]
cvtools.imwrite(
cvtools.draw_boxes_texts(gt_img,
segs,
draw_start=False,
box_format='polygon'),
osp.join(self.opt.debug_dir, 'gt_' + file_name))
bad_num = 0
gt_det = []
for k in range(num_objs):
ann = anns[k]
cls_id = int(self.cat_ids[ann['category_id']])
segm = np.array(ann['segmentation'][0])
# if flipped:
# for i in range(0, len(segm), 2):
# segm[i] = width - segm[i] - 1
# for i in range(0, len(segm), 2):
# segm[i:i + 2] = affine_transform(segm[i:i + 2], trans_output)
# segm[i] = np.clip(segm[i], 0, output_w - 1)
# segm[i + 1] = np.clip(segm[i + 1], 0, output_h - 1)
segm_hull = cv2.convexHull(segm.reshape(-1, 2).astype(np.float32),
clockwise=False)
xy, (w, h), a = cv2.minAreaRect(segm_hull)
if xy[0] > output_w or xy[0] < 0 or xy[1] > output_h or xy[1] < 0:
# TODO：查明为何会出现这种情况。P0750
# xy中y下出现负值或大于127
# print(file_name, ann, segm, xy)
bad_num += 1
continue
if h > 0 and w > 0:
a, w, h = convert_angle(a, w, h, self.opt.a_method)
ct = np.array(xy, dtype=np.float32)
ct_int = ct.astype(np.int32)
radius = gaussian_radius((math.ceil(h), math.ceil(w)))
radius = max(0, int(radius))
radius = self.opt.hm_gauss if self.opt.mse_loss else radius
# radius = np.array((h / 3., w / 3.), np.int32)
draw_gaussian(hm[cls_id], ct_int, radius)
wh[k] = 1. * w, 1. * h
gt_a = a / 90.
if self.opt.a_method == 2:
gt_a = (a + 90.) / 180.
angle[k] = gt_a
ind[k] = ct_int[1] * output_w + ct_int[0]
reg[k] = ct - ct_int
reg_mask[k] = 1
cat_spec_wh[k, cls_id * 2:cls_id * 2 + 2] = wh[k]
cat_spec_mask[k, cls_id * 2:cls_id * 2 + 2] = 1
if self.opt.dense_wh:
draw_dense_reg(dense_wh, hm.max(axis=0), ct_int, wh[k],
radius)
gt_det.append(segm + [cls_id])
else:
bad_num += 1
if bad_num == num_objs: return []
ret = {
'input': inp,
'hm': hm,
'reg_mask': reg_mask,
'ind': ind,
'wh': wh,
'a': angle
}
if self.opt.dense_wh:
hm_a = hm.max(axis=0, keepdims=True)
dense_wh_mask = np.concatenate([hm_a, hm_a], axis=0)
ret.update({'dense_wh': dense_wh, 'dense_wh_mask': dense_wh_mask})
del ret['wh']
elif self.opt.cat_spec_wh:
ret.update({
'cat_spec_wh': cat_spec_wh,
'cat_spec_mask': cat_spec_mask
})
del ret['wh']
if self.opt.reg_offset:
ret.update({'reg': reg})
if self.opt.debug > 0 or 'train' not in self.split:
gt_det = np.array(gt_det, dtype=np.float32) if len(gt_det) > 0 else \
np.zeros((1, 6), dtype=np.float32)
meta = {'c': c, 's': s, 'gt_det': gt_det, 'img_id': img_id}
ret['meta'] = meta
rets.append(ret)
return rets