def make_efficient_example(ex, further_expansion_factor=1, further_scale_up=1, dir_suffix=''): """Make example by storing the image in a cropped and resized version for efficient loading""" # Determine which area we will need from the image # This is a bit larger than the tight crop because of the geometric augmentations max_rotate = np.pi / 6 padding_factor = 1 / 0.85 scale_up_factor = 1 / 0.85 * further_scale_up scale_down_factor = 1 / 0.85 shift_factor = 1.1 base_dst_side = 256 box_center = boxlib.center(ex.bbox) s, c = np.sin(max_rotate), np.cos(max_rotate) w, h = ex.bbox[2:] rot_bbox_side = max(c * w + s * h, c * h + s * w) rot_bbox = boxlib.box_around(box_center, rot_bbox_side) scale_factor = min(base_dst_side / np.max(ex.bbox[2:]) * scale_up_factor, 1) expansion_factor = (padding_factor * shift_factor * scale_down_factor * further_expansion_factor) expanded_bbox = boxlib.expand(rot_bbox, expansion_factor) expanded_bbox = boxlib.intersect(expanded_bbox, np.array([0, 0, 1000, 1000])) new_camera = copy.deepcopy(ex.camera) new_camera.intrinsic_matrix[:2, 2] -= expanded_bbox[:2] new_camera.scale_output(scale_factor) new_camera.undistort() new_im_relpath = ex.image_path.replace('h36m', f'h36m_downscaled{dir_suffix}') new_im_path = f'{paths.DATA_ROOT}/{new_im_relpath}' if not (util.is_file_newer(new_im_path, "2019-11-14T23:33:14") and improc.is_image_readable(new_im_path)): im = improc.imread_jpeg(ex.image_path) dst_shape = improc.rounded_int_tuple(scale_factor * expanded_bbox[[3, 2]]) new_im = cameralib.reproject_image(im, ex.camera, new_camera, dst_shape) util.ensure_path_exists(new_im_path) imageio.imwrite(new_im_path, new_im) new_bbox_topleft = cameralib.reproject_image_points( ex.bbox[:2], ex.camera, new_camera) new_bbox = np.concatenate([new_bbox_topleft, ex.bbox[2:] * scale_factor]) ex = ps3d.Pose3DExample(new_im_relpath, ex.world_coords, new_bbox, new_camera, activity_name=ex.activity_name) return ex
def make_efficient_example(ex, root_muco, i_person): image_relpath = ex.image_path max_rotate = np.pi / 6 padding_factor = 1 / 0.85 scale_up_factor = 1 / 0.85 scale_down_factor = 1 / 0.85 shift_factor = 1.2 base_dst_side = 256 box_center = boxlib.center(ex.bbox) s = np.sin(max_rotate) c = np.cos(max_rotate) rot_bbox_size = (np.array([[c, s], [s, c]]) @ ex.bbox[2:, np.newaxis])[:, 0] side = np.max(rot_bbox_size) rot_bbox_size = np.array([side, side]) rot_bbox = boxlib.box_around(box_center, rot_bbox_size) scale_factor = min(base_dst_side / np.max(ex.bbox[2:]) * scale_up_factor, 1) expansion_factor = padding_factor * shift_factor * scale_down_factor expanded_bbox = boxlib.expand(rot_bbox, expansion_factor) expanded_bbox = boxlib.intersect(expanded_bbox, boxlib.full_box([2048, 2048])) new_camera = ex.camera.copy() new_camera.intrinsic_matrix[:2, 2] -= expanded_bbox[:2] new_camera.scale_output(scale_factor) new_camera.undistort() dst_shape = improc.rounded_int_tuple(scale_factor * expanded_bbox[[3, 2]]) new_im_path = f'{root_muco}_downscaled/{image_relpath[:-4]}_{i_person:01d}.jpg' if not (util.is_file_newer(new_im_path, "2020-02-15T23:28:26")): im = improc.imread_jpeg(f'{root_muco}/{image_relpath}') new_im = cameralib.reproject_image(im, ex.camera, new_camera, dst_shape, antialias_factor=4) util.ensure_path_exists(new_im_path) imageio.imwrite(new_im_path, new_im, quality=95) new_bbox_topleft = cameralib.reproject_image_points(ex.bbox[:2], ex.camera, new_camera) new_bbox = np.concatenate([new_bbox_topleft, ex.bbox[2:] * scale_factor]) if ex.mask is None: noext, ext = os.path.splitext(image_relpath[:-4]) noext = noext.replace('unaugmented_set_001/', '') mask = improc.decode_mask(util.load_pickle(f'{root_muco}/masks/{noext}.pkl')) else: mask = ex.mask if mask is False: new_mask_encoded = None else: new_mask = cameralib.reproject_image(mask, ex.camera, new_camera, dst_shape) new_mask_encoded = improc.encode_mask(new_mask) return p3ds.Pose3DExample( os.path.relpath(new_im_path, paths.DATA_ROOT), ex.world_coords.astype(np.float32), new_bbox.astype(np.float32), new_camera, mask=new_mask_encoded, univ_coords=ex.univ_coords.astype(np.float32))
def get_expanded_crop_box(bbox, full_box, further_expansion_factor): max_rotate = np.pi / 6 padding_factor = 1 / 0.85 scale_down_factor = 1 / 0.85 shift_factor = 1.1 s, c = np.sin(max_rotate), np.cos(max_rotate) w, h = bbox[2:] box_center = boxlib.center(bbox) rot_bbox_side = max(c * w + s * h, c * h + s * w) rot_bbox = boxlib.box_around(box_center, rot_bbox_side) expansion_factor = (padding_factor * shift_factor * scale_down_factor * further_expansion_factor) expanded_bbox = boxlib.intersect(boxlib.expand(rot_bbox, expansion_factor), full_box) return expanded_bbox
def make_efficient_example(ex): image_relpath = ex.image_path max_rotate = np.pi / 6 padding_factor = 1 / 0.85 scale_up_factor = 1 / 0.85 scale_down_factor = 1 / 0.85 shift_factor = 1.2 base_dst_side = 256 box_center = boxlib.center(ex.bbox) s, c = np.sin(max_rotate), np.cos(max_rotate) w, h = ex.bbox[2:] rot_bbox_side = max(c * w + s * h, c * h + s * w) rot_bbox = boxlib.box_around(box_center, rot_bbox_side) scale_factor = min(base_dst_side / np.max(ex.bbox[2:]) * scale_up_factor, 1) expansion_factor = padding_factor * shift_factor * scale_down_factor expanded_bbox = boxlib.expand(rot_bbox, expansion_factor) expanded_bbox = boxlib.intersect(expanded_bbox, np.array([0, 0, 2048, 2048])) new_camera = ex.camera.copy() new_camera.intrinsic_matrix[:2, 2] -= expanded_bbox[:2] new_camera.scale_output(scale_factor) new_camera.undistort() dst_shape = improc.rounded_int_tuple(scale_factor * expanded_bbox[[3, 2]]) new_im_relpath = ex.image_path.replace('3dhp', f'3dhp_downscaled') new_im_path = os.path.join(paths.DATA_ROOT, new_im_relpath) if not (util.is_file_newer(new_im_path, "2019-11-14T23:32:07") and improc.is_image_readable(new_im_path)): im = improc.imread_jpeg(f'{paths.DATA_ROOT}/{image_relpath}') new_im = cameralib.reproject_image(im, ex.camera, new_camera, dst_shape) util.ensure_path_exists(new_im_path) imageio.imwrite(new_im_path, new_im) new_bbox_topleft = cameralib.reproject_image_points( ex.bbox[:2], ex.camera, new_camera) new_bbox = np.concatenate([new_bbox_topleft, ex.bbox[2:] * scale_factor]) mask_rle_relpath = new_im_path.replace('Images', 'FGmaskImages').replace( '.jpg', '.pkl') mask_rle_path = os.path.join(paths.DATA_ROOT, mask_rle_relpath) if util.is_file_newer(mask_rle_path, "2020-03-11T20:46:46"): mask_runlength = util.load_pickle(mask_rle_path) else: mask_relpath = ex.image_path.replace('Images', 'FGmaskImages').replace( '.jpg', '.png') mask = imageio.imread(os.path.join(paths.DATA_ROOT, mask_relpath)) mask_reproj = cameralib.reproject_image(mask, ex.camera, new_camera, dst_shape) mask_runlength = get_mask_with_highest_iou(mask_reproj, new_bbox) util.dump_pickle(mask_runlength, mask_rle_path) return p3ds.Pose3DExample(new_im_relpath, ex.world_coords, new_bbox, new_camera, mask=mask_runlength, univ_coords=ex.univ_coords)