def visualize_box(self, images, targets, h, w, det, img_idx, tb_writer):
det_ = det.cpu().numpy()
# det_ = det_[det_[:, 4] > 0.5]
images = images.permute(0, 2, 3, 1)
images = images.data.cpu().numpy()
for idx in range(len(images)):
img = images[idx].copy()
img = img[:, :, (2, 1, 0)]
img += np.array((104., 117., 123.), dtype=np.float32) # TODO cfg
det__ = det_[det_[:, 5] == idx]
w_ = w[idx, :].cpu().numpy()
h_ = h[idx, :].cpu().numpy()
w_r = 1000 # resize to 1000, h
h_r = w_r / w_ * h_
det__[:, 0:4:2] = det__[:, 0:4:2] / w_ * w_r
det__[:, 1:4:2] = det__[:, 1:4:2] / h_ * h_r
t = targets[idx].numpy() # ground truth
t[:, 0:4:2] = t[:, 0:4:2] * w_r
t[:, 1:4:2] = t[:, 1:4:2] * h_r
t[:, 4] += 1 # TODO because of the traget transformer
boxes = {'gt': t, 'pred': det__}
tb_writer.cfg['steps'] = img_idx + idx
if self.name == 'MS COCO':
tb_writer.cfg['img_id'] = int(
det__[0, 7]) if det__.size != 0 else 'no_detect'
if self.name == 'VOC0712':
tb_writer.cfg['img_id'] = int(
det__[0, 5]) if det__.size != 0 else 'no_detect'
tb_writer.cfg['thresh'] = 0.5
visualize_utils.vis_img_box(img, boxes, (h_r, w_r), tb_writer)
def __call__(self, img, boxes=None, labels=None, tb_writer=None):
if tb_writer is not None:
tb_writer.cfg['aug_name'] = '0_input'
vis_img_box(img, boxes, None, tb_writer)
for i, t in enumerate(self.transforms):
img, boxes, labels = t(img, boxes, labels)
if tb_writer is not None and i + 1 in tb_writer.cfg['vis_list']:
tb_writer.cfg['aug_name'] = '{}_{}'.format(
i + 1,
type(t).__name__)
vis_img_box(img, boxes, None, tb_writer)
return img, boxes, labels
def __call__(self, img, boxes=None, labels=None, tb_writer=None):
if tb_writer is not None:
tb_writer.cfg['aug_name'] = '0_input'
target = np.hstack((boxes, np.expand_dims(labels + 1, axis=1)))
vis_img_box(img, target, None, tb_writer)
for i, t in enumerate(self.transforms):
img, boxes, labels = t(img, boxes, labels)
if boxes is None: continue
target = np.hstack((boxes, np.expand_dims(labels + 1, axis=1)))
if tb_writer is not None and i + 1 in tb_writer.cfg['aug_vis_list']:
tb_writer.cfg['aug_name'] = '{}_{}'.format(
i + 1,
type(t).__name__)
vis_img_box(img, target, None, tb_writer)
return img, boxes, labels
def visualize_box(self, images, targets, h, w, det, img_idx, tb_writer):
det_ = det.cpu().numpy()
# det_ = det_[det_[:, 4] > 0.5]
images = images.permute(0, 2, 3, 1) #N H W C
images = images.data.cpu().numpy()
#print(det_.shape, 'ssss')
for idx in range(len(images)):
img = images[idx].copy()
img += np.array((104., 117., 123.), dtype=np.float32)
w_r = 300 # resize to 1000, h
h_r = 300
boxes = None
det__ = np.array([])
if det_.size != 0:
det__ = det_[det_[:, 5] == idx]
w_ = w[idx, :].cpu().numpy()
h_ = h[idx, :].cpu().numpy()
# w_r = 1000 # resize to 1000, h
# h_r = w_r / w_ * h_
det__[:, 0:4:2] = det__[:, 0:4:2] / w_ * w_r
det__[:, 1:4:2] = det__[:, 1:4:2] / h_ * h_r
if tb_writer.cfg['phase'] == 'train':
w_ = 1.
h_ = 1.
t = targets[idx].clone().numpy() # ground truth
t[:, 0:4:2] = t[:, 0:4:2] / w_ * w_r
t[:, 1:4:2] = t[:, 1:4:2] / h_ * h_r
t[:, 4] += 1 # TODO because of the traget transformer #label
boxes = {'gt': t, 'pred': det__} #Shape (*, 5) (*, 7)
tb_writer.cfg['steps'] = img_idx + idx
if self.name == 'MS COCO':
tb_writer.cfg['img_id'] = int(
det__[0, 7]) if det__.size != 0 else 'no_detect'
if self.name == 'VOC0712' or 'TEMP_LP':
tb_writer.cfg['img_id'] = int(
det__[0, 5]) if det__.size != 0 else 'no_detect'
tb_writer.cfg['thresh'] = 0.3
visualize_utils.vis_img_box(img, boxes, (h_r, w_r),
tb_writer) #(h_r, w_r)