def get_mask(i_subject, i_seq, i_cam, i_frame):
chroma_frame = improc.imread_jpeg(
f'{paths.DATA_ROOT}/3dhp/S{i_subject}/Seq{i_seq}/FGmasks/img_{i_cam}_{i_frame:06d}.jpg'
)
person_box = get_box(i_subject, i_seq, i_cam, i_frame)
is_fg = chroma_frame[..., 0] > 100
n_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(
is_fg.astype(np.uint8), 4, cv2.CV_32S)
component_boxes = stats[:, :4]
ious = [
boxlib.iou(component_box, person_box)
for component_box in component_boxes
]
ious[0] = 0
person_label = np.argmax(ious)
mask = (labels == person_label).astype(np.uint8)
# Remove foreground pixels that are far from the person box
intbox = boxlib.intersect(boxlib.full_box((2048, 2048)),
boxlib.expand(person_box, 1.3)).astype(int)
mask[:intbox[1]] = 0
mask[:, :intbox[0]] = 0
mask[:, intbox[0] + intbox[2]:] = 0
mask[intbox[1] + intbox[3]:] = 0
encoded_mask = improc.encode_mask(mask)
return encoded_mask
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, rect_id):
"""Make example by storing the image in a cropped and resized version for efficient loading"""
# Determine which area we will need
# For rotation, usual padding around box, scale (shrink) augmentation and shifting
padding_factor = 1 / 0.85
scale_up_factor = 1 / 0.85
scale_down_factor = 1 / 0.85
shift_factor = 1.1
max_rotate = np.pi / 6
rot_factor = np.sin(max_rotate) + np.cos(max_rotate)
base_dst_side = 256
scale_factor = min(base_dst_side / ex.bbox[3] * scale_up_factor, 1)
hopeful_factor = 0.9
expansion_factor = (
rot_factor * padding_factor * shift_factor * scale_down_factor * hopeful_factor)
expanded_bbox = boxlib.expand(boxlib.expand_to_square(ex.bbox), expansion_factor)
imsize = improc.image_extents(ex.image_path)
full_box = np.array([0, 0, imsize[0], imsize[1]])
expanded_bbox = boxlib.intersect(expanded_bbox, full_box)
old_camera = cameralib.Camera.create2D()
new_camera = old_camera.copy()
new_camera.shift_image(-expanded_bbox[:2])
new_camera.scale_output(scale_factor)
dst_shape = improc.rounded_int_tuple(scale_factor * expanded_bbox[[3, 2]])
new_im_path = ex.image_path.replace('mpii', f'mpii_downscaled')
without_ext, ext = os.path.splitext(new_im_path)
new_im_path = f'{without_ext}_{rect_id:02d}{ext}'
if not (util.is_file_newer(new_im_path, "2019-11-12T17:54:06") and
improc.is_image_readable(new_im_path)):
im = improc.imread_jpeg(ex.image_path)
new_im = cameralib.reproject_image(im, old_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], old_camera, new_camera)
new_bbox = np.concatenate([new_bbox_topleft, ex.bbox[2:] * scale_factor])
new_coords = cameralib.reproject_image_points(ex.coords, old_camera, new_camera)
ex = Pose2DExample(os.path.relpath(new_im_path, paths.DATA_ROOT), new_coords, bbox=new_bbox)
return ex
def make_muco():
joint_info, selected_joints = make_joint_info()
root_3dhp = f'{paths.DATA_ROOT}/3dhp'
root_muco = f'{paths.DATA_ROOT}/muco'
sample_info = np.load(f'{root_muco}/composite_frame_origins.npy')
n_all_joints = 28
valid_indices = list(np.load(f'{root_muco}/valid_composite_frame_indices.npy'))
all_detections = util.load_pickle(f'{root_muco}/yolov3_detections.pkl')
all_detections = np.array([all_detections[k] for k in sorted(all_detections.keys())])
all_visible_boxes = np.load(f'{root_muco}/visible_boxes.npy')
matloader = functools.lru_cache(1024)(matlabfile.load)
@functools.lru_cache(1024)
def get_world_coords(i_subject, i_seq, i_cam, anno_name):
seqpath = f'{root_3dhp}/S{i_subject}/Seq{i_seq}'
anno_file = matloader(f'{seqpath}/annot.mat')
camcoords = anno_file[anno_name][i_cam].reshape(
[-1, n_all_joints, 3])[:, selected_joints]
camera = load_cameras(f'{seqpath}/camera.calibration')[i_cam]
world_coords = [camera.camera_to_world(c) for c in camcoords]
return world_coords
examples = []
with util.BoundedPool(None, 120) as pool:
for i_sample, people, detections, visible_boxes in zip(
util.progressbar(valid_indices), sample_info[valid_indices],
all_detections[valid_indices], all_visible_boxes[valid_indices]):
detections = [box for box in detections if box[-1] > 0.1]
if not detections:
continue
filename = f'{i_sample + 1:06d}.jpg'
image_relpath = f'unaugmented_set_001/{filename[:2]}/{filename[:4]}/{filename}'
gt_people = []
for i_person, ((i_subject, i_seq, i_cam, i_frame), visible_box) in enumerate(
zip(people, visible_boxes)):
seqpath = f'{root_3dhp}/S{i_subject}/Seq{i_seq}'
world_coords = get_world_coords(i_subject, i_seq, i_cam, 'annot3')[i_frame]
univ_world_coords = get_world_coords(
i_subject, i_seq, i_cam, 'univ_annot3')[i_frame]
camera = load_cameras(f'{seqpath}/camera.calibration')[i_cam]
im_coords = camera.world_to_image(world_coords)
coord_bbox = boxlib.expand(boxlib.intersect(
boxlib.bb_of_points(im_coords),
boxlib.full_box([2048, 2048])), 1.05)
bbox = boxlib.intersect_vertical(visible_box, coord_bbox)
ex = p3ds.Pose3DExample(
image_relpath, world_coords, bbox, camera, mask=None,
univ_coords=univ_world_coords)
gt_people.append(ex)
if not gt_people:
continue
iou_matrix = np.array([[boxlib.iou(gt_person.bbox, box[:4])
for box in detections]
for gt_person in gt_people])
gt_indices, det_indices = scipy.optimize.linear_sum_assignment(-iou_matrix)
for i_gt, i_det in zip(gt_indices, det_indices):
gt_box = gt_people[i_gt].bbox
det_box = detections[i_det]
if (iou_matrix[i_gt, i_det] > 0.1 and
boxlib.area(det_box) < 2 * boxlib.area(gt_box)):
ex = gt_people[i_gt]
ex.bbox = np.array(detections[i_det][:4])
pool.apply_async(make_efficient_example, (ex, root_muco, i_gt),
callback=examples.append)
examples.sort(key=lambda ex: ex.image_path)
return p3ds.Pose3DDataset(joint_info, examples)
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)
def get_box(i_subject, i_seq, i_cam, i_frame):
imcoords = get_coords(i_subject, i_seq, i_cam, 'annot3')[2][i_frame]
box = boxlib.expand(boxlib.bb_of_points(imcoords), 1.05)
return boxlib.intersect(boxlib.full_box((2048, 2048)),
box).astype(np.float32)
