label_path_for_show[b] = os.path.basename(cur_data_batch[1][b])
cur_img = cv2.imread(cur_data_batch[0][b])
cur_label = np.load(cur_data_batch[1][b]).tolist()
cur_joints_zidx = (
cur_label["joints_zidx"] -
1).copy() # cause lua is from 1 to n not 0 to n-1
cur_joints = np.concatenate(
[cur_label["joints_2d"], cur_joints_zidx[:, np.newaxis]],
axis=1)
cur_img, cur_joints = preprocessor.preprocess(
cur_img,
cur_joints,
is_training=not is_valid,
is_rotate=False)
# generate the heatmaps and volumes
batch_images_np[b] = cur_img
hm_joint_2d = np.round(cur_joints[:, 0:2] /
configs.coords_2d_scale)
hm_joint_3d = np.concatenate(
[hm_joint_2d, cur_joints[:, 2][:, np.newaxis]], axis=1)
batch_centers_np[b] = hm_joint_3d
acc_hm = 0
acc_vol = 0
if is_valid:
cur_img = cv2.imread(img_list[data_index.val])
cur_label = np.load(lbl_list[data_index.val]).tolist()
data_index.val += 1
########## Save the data for evaluation ###########
source_txt_arr.append(cur_label["source"])
center_arr.append(cur_label["center"])
scale_arr.append(cur_label["scale"])
depth_root_arr.append(cur_label["joints_3d"][0, 2])
gt_joints_3d_arr.append(cur_label["joints_3d"].copy())
crop_joints_2d_arr.append(cur_label["joints_2d"].copy())
cam_matrix_arr.append(cur_label["cam_mat"].copy())
###################################################
cur_img, _ = preprocessor.preprocess(cur_img,
None,
is_training=False)
batch_images_np[b] = cur_img
batch_images_flipped_np[b] = preprocessor.flip_img(
batch_images_np[b])
mean_vol_joints, \
raw_vol_joints = sess.run(
[
ordinal_model.mean_joints,
ordinal_model.raw_joints
],
feed_dict={input_images: np.concatenate([batch_images_np, batch_images_flipped_np], axis=0)})
print((len(img_path_for_show) * "{}\n").format(
cur_label = np.load(
scale_lbl_list[scale_data_index.val]).tolist()
scale_data_index.val += 1
########## Save the data for evaluation ###########
gt_joints_3d = cur_label["joints_3d"].copy()
gt_depth_arr.append(gt_joints_3d[:, 2] -
gt_joints_3d[0, 2])
###################################################
cur_joints = np.concatenate([
cur_label["joints_2d"],
cur_label["joints_3d"][:, 2][:, np.newaxis]
],
axis=1)
cur_img, cur_joints = preprocessor.preprocess(
cur_img, cur_joints, is_training=False)
batch_images_np[b] = cur_img
batch_images_flipped_np[b] = preprocessor.flip_img(
batch_images_np[b])
mean_scale_depth,\
raw_scale_depth = sess.run(
[
ordinal_model.mean_volumes_z,
ordinal_model.raw_volumes_z
],
feed_dict={input_images: np.concatenate([batch_images_np, batch_images_flipped_np], axis=0)})
scale_for_show = []
scale_depth = mean_scale_depth