def _crf_with_alpha(ori_img,cam_dict, alpha):
v = np.array(list(cam_dict.values()))
bg_score = np.power(1 - np.max(v, axis=0, keepdims=True), alpha)
bgcam_score = np.concatenate((bg_score, v), axis=0)
crf_score = imutils.crf_inference(ori_img, bgcam_score, labels=bgcam_score.shape[0])
n_crf_al = np.zeros([21, bg_score.shape[1], bg_score.shape[2]])
n_crf_al[0, :, :] = crf_score[0, :, :]
for i, key in enumerate(cam_dict.keys()):
n_crf_al[key + 1] = crf_score[i + 1]
return n_crf_al
def _crf_with_alpha(cam_dict, alpha):
v = np.array(list(cam_dict.values()))
bg_score = np.power(1 - np.max(v, axis=0, keepdims=True), alpha)
bgcam_score = np.concatenate((bg_score, v), axis=0)
print("v.shape ", v.shape)
print("orig_img.shape ", orig_img.shape)
print("type(bgcam_score)", type(bgcam_score))
print("type(orig_img) ", type(orig_img))
crf_score = imutils.crf_inference(orig_img,
bgcam_score,
labels=bgcam_score.shape[0])
print("after CRF......")
input("[130].......")
n_crf_al = dict()
n_crf_al[0] = crf_score[0]
for i, key in enumerate(cam_dict.keys()):
n_crf_al[key + 1] = crf_score[i + 1]
return n_crf_al
def rw_crf(predicted_dict, name=None):
"""
- orig_img: [3,H,W], np array
- predicted_dict: dictionary, each item is a [H,W] predicted score for coresponding class
"""
v = np.array(list(predicted_dict.values()))
pred_softmax = torch.nn.Softmax(dim=0)
probs = pred_softmax(torch.tensor(v)).numpy()
# orig_img = orig_img.data.cpu().numpy().transpose((1, 2, 0)) # [H,W,3]
img_path = voc12.data.get_img_path(name, args.voc12_root)
orig_img = np.asarray(Image.open(img_path))
# === note that orig_img have the shape [H,W,3]
crf_score = imutils.crf_inference(orig_img,
probs,
labels=v.shape[0])
h, w = orig_img.shape[:2]
crf_score_np = np.zeros(shape=(args.num_class, h, w))
crf_dict = dict()
for i, key in enumerate(predicted_dict.keys()):
crf_score_np[key] = crf_score[i]
crf_dict[key] = crf_score[i]
return crf_score_np, crf_dict
cam_full_arr = torch.from_numpy(cam_full_arr)
cam_full_arr = F.avg_pool2d(cam_full_arr, 8, 8)
cam_vec = cam_full_arr.view(21, -1)
cam_rw = torch.matmul(cam_vec.cuda(), trans_mat)
cam_rw = cam_rw.view(1, 21, dheight, dwidth)
cam_rw = torch.nn.Upsample((img.shape[2], img.shape[3]),
mode='bilinear')(cam_rw)
cam_rw = cam_rw[:, :, :orig_shape[2], :orig_shape[3]]
cam_rw = cam_rw.squeeze().data.cpu().numpy()
aff_dict = {}
aff_dict[0] = cam_rw[0]
for i in range(20):
if label[i] > 1e-5:
aff_dict[i + 1] = cam_rw[i + 1]
np.save(os.path.join(save_dir_aff, name + '.npy'), aff_dict)
mask = np.argmax(cam_rw, axis=0)
mask2png(mask, os.path.join(save_dir_aff, name + '.png'))
v = np.array(list(aff_dict.values()))
aff_crf = imutils.crf_inference(orig_img, v, labels=v.shape[0])
aff_crf_full_arr = np.zeros((21, orig_shape[2], orig_shape[3]),
np.float32)
cnt = 0
for k, v in aff_dict.items():
aff_crf_full_arr[k] = aff_crf[cnt]
cnt += 1
np.save(os.path.join(save_dir_aff_crf, name + '.npy'), aff_crf)
mask = np.argmax(aff_crf_full_arr, axis=0)
mask2png(mask, os.path.join(save_dir_aff_crf, name + '.png'))
cam_vec = cam_full_arr.view(21, -1)
cam_rw = torch.matmul(cam_vec.cuda(), trans_mat)
cam_rw = cam_rw.view(1, 21, dheight, dwidth)
cam_rw = torch.nn.Upsample((img.shape[2], img.shape[3]),
mode='bilinear')(cam_rw)
if args.crf:
img_8 = img[0].numpy().transpose((1, 2, 0))
img_8 = np.ascontiguousarray(img_8)
mean = (0.485, 0.456, 0.406)
std = (0.229, 0.224, 0.225)
img_8[:, :, 0] = (img_8[:, :, 0] * std[0] + mean[0]) * 255
img_8[:, :, 1] = (img_8[:, :, 1] * std[1] + mean[1]) * 255
img_8[:, :, 2] = (img_8[:, :, 2] * std[2] + mean[2]) * 255
img_8[img_8 > 255] = 255
img_8[img_8 < 0] = 0
img_8 = img_8.astype(np.uint8)
cam_rw = cam_rw[0].cpu().numpy()
cam_rw = imutils.crf_inference(img_8, cam_rw, t=1)
cam_rw = torch.from_numpy(cam_rw).view(1, 21, img.shape[2],
img.shape[3]).cuda()
_, cam_rw_pred = torch.max(cam_rw, 1)
res = np.uint8(
cam_rw_pred.cpu().data[0])[:orig_shape[2], :orig_shape[3]]
scipy.misc.imsave(os.path.join(args.out_rw, name + '.png'), res)
