def export_scenelet(um, o_pos_3d, o_polys_3d, query_full_skeleton, scenes, joints_active, transform_id=None): """Extract a scenelet (poses and objects) from the data from the optimized problem. Args: um (stealth.pose.unk_manager.UnkManager): Data manager. o_pos_3d (np.ndarray): Output 3D poses. o_polys_3d (np.ndarray): (6K, 4, 3) 3D oriented bounding boxes stored stacked. query_full_skeleton (stealth.logic.skeleton.Skeleton): Initial path containing time information. joints_active (list): List of joint_ids that were optimized for. Usage: pose16[:, joints_active] = o_pos_3d[pid, :, :] transform_id (int): Export only a specific group. Everything is exported, if None. Returns: A scenelet extracted from the data provided. """ # cache function _guess_time_at = query_full_skeleton.guess_time_at # all poses or the ones that belong to a group/scenelet if transform_id is None: pids_sorted = sorted([(pid, pid2scene) for pid, pid2scene in um.pids_2_scenes.items()], key=lambda e: e[1].frame_id) else: # pids_sorted = sorted([(pid, pid2scene) # for pid, pid2scene in um.pids_2_scenes.items() # if pid2scene.transform_id == transform_id], # key=lambda e: e[1].frame_id) pids_2_scenes = um.pids_2_scenes pids_sorted = sorted([(pid, pids_2_scenes[pid]) for pid in um.get_pids_for(transform_id)], key=lambda e: e[1].frame_id) # create output scenelet o = Scenelet() charness = None # # Skeleton # # cache skeleton reference skeleton = o.skeleton # fill skeleton for pid, pid2scene in pids_sorted: if charness is None: scene = scenes[pid2scene.id_scene] charness = scene.charness o.add_aux_info('name_scenelet', scene.name_scenelet) o.charness = charness # get frame_id frame_id = int(pid2scene.frame_id) # check if already exists if skeleton.has_pose(frame_id): # TODO: fix overlapping frame_ids lg.warning("[export_scenelet] Overwriting output frame_id %d" % frame_id) # add with time guessed from input skeleton rate pose = np.zeros((3, Joint.get_num_joints())) pose[:, joints_active] = o_pos_3d[pid, :, :] pose[:, Joint.PELV] = (pose[:, Joint.LHIP] + pose[:, Joint.RHIP]) / 2. pose[:, Joint.NECK] = (pose[:, Joint.HEAD] + pose[:, Joint.THRX]) / 2. # for j, jid in joints_remap.items(): # pose[:, j] = o_pos_3d[pid, :, jid] assert not skeleton.has_pose(frame_id=frame_id), \ 'Already has pose: {}'.format(frame_id) skeleton.set_pose(frame_id=frame_id, pose=pose, time=_guess_time_at(frame_id)) # # Objects # scene_obj = None scene_obj_oid = 0 # unique identifier that groups parts to objects for polys2scene in um.polys2scene.values(): # Check, if we are restricted to a certain group if transform_id is not None \ and polys2scene.transform_id != transform_id: continue start = polys2scene.poly_id_start end = start + polys2scene.n_polys # 6 x 4 x 3 polys = o_polys_3d[start:end, ...] assert polys.shape[0] == 6, "Assumed cuboids here" if scene_obj is None or scene_obj_oid != polys2scene.object_id: category = next(cat for cat in CATEGORIES if CATEGORIES[cat] == polys2scene.cat_id) scene_obj = SceneObj(label=category) scene_obj_oid = polys2scene.object_id o.add_object(obj_id=-1, scene_obj=scene_obj, clone=False) part = scene_obj.add_part(part_id=-1, label_or_part=polys2scene.part_label) # TODO: average for numerical precision errors centroid = np.mean(polys, axis=(0, 1)) ax0 = polys[0, 1, :] - polys[0, 0, :] scale0 = np.linalg.norm(ax0) ax0 /= scale0 ax1 = polys[0, 3, :] - polys[0, 0, :] scale1 = np.linalg.norm(ax1) ax1 /= scale1 ax2 = polys[1, 0, :] - polys[0, 0, :] scale2 = np.linalg.norm(ax2) ax2 /= scale2 part.obb = Obb(centroid=centroid, axes=np.concatenate( (ax0[:, None], ax1[:, None], ax2[:, None]), axis=1), scales=[scale0, scale1, scale2]) # if scene_obj is not None: # o.add_object(obj_id=-1, scene_obj=scene_obj, clone=False) # else: # lg.warning("No objects in scenelet?") # scene_obj = SceneObj('couch') # for poly_id in range(0, o_polys_3d.shape[0], 6): # rects = o_polys_3d[poly_id : poly_id + 6, ...] # # lg.debug("rects:\n%s" % rects) # scene_obj.add_part(poly_id, 'seat') # # # fig = plt.figure() # # ax = fig.add_subplot(111, projection='3d') # # for rid, rect in enumerate(rects): # # wrapped = np.concatenate((rect, rect[0:1, :]), axis=0) # # ax.plot(wrapped[:, 0], wrapped[:, 2], wrapped[:, 1]) # # for ci in range(4): # # c = rect[ci, :] # # ax.text(c[0], c[2], c[1], s="%d, %d, %d" # # % (poly_id, rid, ci)) # # if rid >= 1: # # break # # # # plt.show() # part = scene_obj.get_part(poly_id) # centroid = np.mean(rects, axis=(0, 1)) # ax0 = rects[0, 1, :] - rects[0, 0, :] # scale0 = np.linalg.norm(ax0) # ax0 /= scale0 # ax1 = rects[0, 3, :] - rects[0, 0, :] # scale1 = np.linalg.norm(ax1) # ax1 /= scale1 # ax2 = rects[1, 0, :] - rects[0, 0, :] # scale2 = np.linalg.norm(ax2) # ax2 /= scale2 # part.obb = Obb(centroid=centroid, # axes=np.concatenate(( # ax0[:, None], ax1[:, None], ax2[:, None] # ), axis=1), # scales=[scale0, scale1, scale2]) # o.add_object(obj_id=99, scene_obj=scene_obj, # clone=False) return o
def main(argv): conf = Conf.get() parser = argparse.ArgumentParser("Denis pose converter") parser.add_argument('camera_name', help="Camera name ('G15', 'S6')", type=str) parser.add_argument( '-d', dest='dir', required=True, help="Path to the <scene folder>/denis containing skeletons.json") parser.add_argument( '-filter', dest='with_filtering', action="store_true", help="Should we do post-filtering (1-euro) on the pelvis positions") parser.add_argument('-huber', required=False, help="Should we do huber loss?", action='store_true') parser.add_argument('-smooth', type=float, default=0.005, help="Should we have a smoothness term (l2/huber)?") parser.add_argument( '--winsorize-limit', type=float, default=conf.optimize_path.winsorize_limit, help='Threshold for filtering too large jumps of the 2D centroid') parser.add_argument('--no-resample', action='store_true', help="add resampled frames") parser.add_argument('--n-actors', type=int, default=1, help="How many skeletons to track.") parser.add_argument('-n-actors', type=int, default=1, help="Max number of people in scene.") # parser.add_argument( # '-r', type=float, # help='Video rate. Default: 1, if avconv -r 5. ' # 'Original video sampling rate (no subsampling) should be ' # '24/5=4.8. avconv -r 10 leads to 24/10=2.4.', # required=True) parser.add_argument('--person_height', type=float, help='Assumed height of human(s) in video.', default=Conf.get().optimize_path.person_height) parser.add_argument( '--forwards-window-size', type=int, help='How many poses in time to look before AND after to ' 'average forward direction. 0 means no averaging. Default: 0.', default=0) parser.add_argument('--no-img', action='store_true', help='Read and write images (vis reproj error)') parser.add_argument('--postfix', type=str, help="output file postfix.", default='unannot') args = parser.parse_args(argv) show = False args.resample = not args.no_resample # assert not args.resample, "resample should be off" assert os.path.exists(args.dir), "Source does not exist: %s" % args.dir p_scene = os.path.normpath(os.path.join(args.dir, os.pardir)) # type: str p_video_params = os.path.join(p_scene, 'video_params.json') assert os.path.exists(p_video_params), "Need video_params.json for rate" if 'r' not in args or args.r is None: args.r = json.load(open(p_video_params, 'r'))['rate-avconv'] # manual parameters (depth initialization, number of actors) p_scene_params = os.path.join(args.dir, os.pardir, 'scene_params.json') if not os.path.exists(p_scene_params): scene_params = { 'depth_init': 10., 'actors': args.n_actors, 'ground_rot': [0., 0., 0.] } json.dump(scene_params, open(p_scene_params, 'w')) raise RuntimeError("Inited scene_params.json, please check: %s" % p_scene_params) else: scene_params = json.load(open(p_scene_params, 'r')) lg.warning("Will work with %d actors and init depth to %g" % (scene_params['actors'], scene_params['depth_init'])) assert '--n-actors' not in argv \ or args.n_actors == scene_params['actors'], \ "Actor count mismatch, remove %d from args, because " \ "scene_params.json says %d?" \ % (args.n_actors, scene_params['actors']) args.n_actors = scene_params['actors'] ground_rot = scene_params['ground_rot'] or [0., 0., 0.] # load images path_images = os.path.abspath(os.path.join(args.dir, os.pardir, 'origjpg')) images = {} shape_orig = None if not args.no_img: images, shape_orig = load_images(path_images) path_skeleton = \ max((f for f in os.listdir(os.path.join(args.dir)) if f.startswith('skeletons') and f.endswith('json')), key=lambda s: int(os.path.splitext(s)[0].split('_')[1])) print("path_skeleton: %s" % path_skeleton) data = json.load(open(os.path.join(args.dir, path_skeleton), 'r')) # data, pose_constraints, first_run = \ # cleanup(data, p_dir=os.path.join(args.dir, os.pardir)) # poses_2d = [] # plt.figure() # show_images(images, data) if False: # pose_ids = identify_actors_multi(data, n_actors=1) p_segm_pickle = os.path.join(args.dir, os.pardir, "label_skeletons.pickle") problem = None if False and os.path.exists(p_segm_pickle): lg.warning("Loading skeleton segmentation from pickle %s" % p_segm_pickle) pose_ids, problem = pickle_load(open(p_segm_pickle, 'rb')) if not problem or problem._n_actors != args.n_actors: pose_ids, problem, data = more_actors_gurobi( data, n_actors=args.n_actors, constraints=pose_constraints, first_run=first_run) if True or show: show_multi(images, data, pose_ids, problem, p_dir=os.path.join(args.dir, os.pardir), first_run=first_run, n_actors=args.n_actors) pickle.dump((pose_ids, problem), open(p_segm_pickle, 'wb'), -1) else: pose_ids = greedy_actors(data, n_actors=args.n_actors) data = DataPosesWrapper(data=data) visible_f = {a: {} for a in range(args.n_actors)} visible_f_max = 0. if show: plt.ion() fig = None axe = None scatters = dict() # how many images we have min_frame_id = min(f for f in pose_ids) frames_mod = max(f for f in pose_ids) - min_frame_id + 1 skel_ours = Skeleton(frames_mod=frames_mod, n_actors=args.n_actors, min_frame_id=min_frame_id) skel_ours_2d = Skeleton(frames_mod=frames_mod, n_actors=args.n_actors, min_frame_id=min_frame_id) # assert len(images) == 0 or max(f for f in images) + 1 == frames_mod, \ # "Assumed image count is %d, but max_frame_id is %d" \ # % (len(images), frames_mod-1) if isinstance(data, DataPosesWrapper): frames = data.get_frames() else: frames = [] for frame_str in sorted(data.get_frames()): try: frame_id = int(frame_str.split('_')[1]) except ValueError: print("skipping key %s" % frame_id) continue frames.append(frame_id) my_visibilities = [[], []] for frame_id in frames: frame_str = DataPosesWrapper._to_frame_str(frame_id) pose_in = data.get_poses_3d(frame_id=frame_id) # np.asarray(data[frame_str][u'centered_3d']) # pose_in_2d = np.asarray(data[frame_str][u'pose_2d']) pose_in_2d = data.get_poses_2d(frame_id=frame_id) # visible = np.asarray(data[frame_str][u'visible']) if False and len(pose_in.shape) > 2: pose_id = pose_ids[frame_id] if not args.no_img: im = cv2.cvtColor(images[frame_id], cv2.COLOR_RGB2BGR) for i in range(pose_in.shape[0]): c = (1., 0., 0., 1.) if i == pose_id: c = (0., 1., 0., 1.) color = tuple(int(c_ * 255) for c_ in c[:3]) for p2d in pose_in_2d[i, :, :]: # color = (c[0] * 255, c[1] * 255., c[2] * 255.) cv2.circle(im, (p2d[1], p2d[0]), radius=3, color=color, thickness=-1) center = np.mean(pose_in_2d[i, :, :], axis=0).round().astype('i4').tolist() cv2.putText(im, "%d" % i, (center[1], center[0]), 1, 1, color) if show: cv2.imshow("im", im) cv2.waitKey(100) # if sid not in scatters: # scatters[sid] = axe.scatter(pose_in_2d[i, :, 1], pose_in_2d[i, :, 0], c=c) # else: # scatters[sid].set_offsets(pose_in_2d[i, :, [1, 0]]) # scatters[sid].set_array(np.tile(np.array(c), pose_in_2d.shape[1])) # scatter.set_color(c) # plt.draw() # plt.pause(1.) pose_in = pose_in[pose_id, :, :] pose_in_2d = pose_in_2d[pose_id, :, :] visible = visible[pose_id] # else: # pose_id = 0 # pose_id = pose_ids[frame_id] for actor_id in range(args.n_actors): # if actor_id in (2, 3, 4, 5, 8, 9) # expanded frame_id frame_id2 = Skeleton.unmod_frame_id(frame_id=frame_id, actor_id=actor_id, frames_mod=frames_mod) assert (actor_id != 0) ^ (frame_id2 == frame_id), "no" frame_id_mod = skel_ours.mod_frame_id(frame_id=frame_id2) assert frame_id_mod == frame_id, \ "No: %d %d %d" % (frame_id, frame_id2, frame_id_mod) actor_id2 = skel_ours.get_actor_id(frame_id2) assert actor_id2 == actor_id, "no: %s %s" % (actor_id, actor_id2) # which pose explains this actor in this frame pose_id = pose_ids[frame_id][actor_id] # check, if actor found if pose_id < 0: continue # 3D pose pose = pose_in[pose_id, :, JointDenis.revmap].T # added by Aron on 4/4/2018 (Denis' pelvis is too high up) pose[:, Joint.PELV] = (pose[:, Joint.LHIP] + pose[:, Joint.RHIP]) \ / 2. skel_ours.set_pose(frame_id2, pose) # 2D pose pose_2d = pose_in_2d[pose_id, :, :] arr = np.array(JointDenis.pose_2d_to_ours(pose_2d), dtype=np.float32).T skel_ours_2d.set_pose(frame_id2, arr) # # visibility (binary) and confidence (float) # # np.asarray(data[frame_str][u'visible'][pose_id]) vis_i = data.get_visibilities(frame_id)[pose_id] # vis_f = np.asarray(data[frame_str][u'visible_float'][pose_id]) vis_f = data.get_confidences(frame_id)[pose_id] for jid, visible in enumerate(vis_i): # for each joint # binary visibility jid_ours = JointDenis.to_ours_2d(jid) skel_ours_2d.set_visible(frame_id2, jid_ours, visible) # confidence (fractional visibility) if np.isnan(vis_f[jid]): continue try: visible_f[actor_id][frame_id2][jid_ours] = vis_f[jid] except KeyError: visible_f[actor_id][frame_id2] = {jid_ours: vis_f[jid]} visible_f_max = max(visible_f_max, vis_f[jid]) conf_ = get_conf_thresholded(vis_f[jid], thresh_log_conf=None, dtype_np=np.float32) skel_ours_2d.set_confidence(frame_id=frame_id2, joint=jid_ours, confidence=conf_) my_visibilities[0].append(vis_f[jid]) my_visibilities[1].append(conf_) skel_ours_2d._confidence_normalized = True plt.figure() plt.plot(my_visibilities[0], my_visibilities[1], 'o') plt.savefig('confidences.pdf') assert skel_ours.n_actors == args.n_actors, "no" assert skel_ours_2d.n_actors == args.n_actors, "no" # align to room min_z = np.min(skel_ours.poses[:, 2, :]) print("min_max: %s, %s" % (min_z, np.max(skel_ours.poses[:, 2, :]))) skel_ours.poses[:, 2, :] += min_z skel_ours.poses /= 1000. # The output is scaled to 2m by Denis. # We change this to 1.8 * a scale in order to correct for # the skeletons being a bit too high still. skel_ours.poses *= \ args.person_height * conf.optimize_path.height_correction / 2. skel_ours.poses[:, 2, :] *= -1. skel_ours.poses = skel_ours.poses[:, [0, 2, 1], :] # refine name_video = args.dir.split(os.sep)[-2] out_path = os.path.join(args.dir, os.pardir, "skel_%s_%s.json" % (name_video, args.postfix)) out_path_orig = os.path.join(args.dir, os.pardir, "skel_%s_lfd_orig.json" % name_video) sclt_orig = Scenelet(skeleton=copy.deepcopy(skel_ours)) sclt_orig.save(out_path_orig) skel_ours_2d_all = copy.deepcopy(skel_ours_2d) assert len(skel_ours_2d_all.get_frames()), skel_ours_2d_all.get_frames() # # Optimize # # frames_ignore = [(282, 372), (516, 1000)] skel_ours, skel_ours_2d, intrinsics, \ frame_ids_filled_in = prepare( args.camera_name, winsorize_limit=args.winsorize_limit, shape_orig=shape_orig, path_scene=p_scene, skel_ours_2d=skel_ours_2d, skel_ours=skel_ours, resample=args.resample, path_skel=path_skeleton) frames_ignore = [] tr_ground = np.eye(4, dtype=np.float32) skel_opt, out_images, K = \ optimize_path( skel_ours, skel_ours_2d, images, intrinsics=intrinsics, path_skel=out_path, shape_orig=shape_orig, use_huber=args.huber, weight_smooth=args.smooth, frames_ignore=frames_ignore, resample=args.resample, depth_init=scene_params['depth_init'], ground_rot=ground_rot) for frame_id in skel_opt.get_frames(): skel_opt.set_time(frame_id=frame_id, time=float(frame_id) / args.r) skel_opt_raw = copy.deepcopy(skel_opt) skel_opt_resampled = Skeleton.resample(skel_opt) # Filter pelvis if args.with_filtering: out_filter_path = os.path.join(args.dir, os.pardir, "vis_filtering") skel_opt = filter_(skel_opt_resampled, out_filter_path=out_filter_path, skel_orig=skel_opt, weight_smooth=args.smooth, forwards_window_size=args.forwards_window_size) else: skel_opt.estimate_forwards(k=args.forwards_window_size) skel_opt_resampled.estimate_forwards(k=args.forwards_window_size) # if len(images): # skel_opt.fill_with_closest(images.keys()[0], images.keys()[-1]) min_y, max_y = skel_opt.get_min_y(tr_ground) print("min_y: %s, max_y: %s" % (min_y, max_y)) # # save # frame_ids_old = set(skel_opt.get_frames()) if args.resample: skel_opt = skel_opt_resampled frame_ids_filled_in.update( set(skel_opt.get_frames()).difference(frame_ids_old)) lg.warning("Saving resampled scenelet!") scenelet = Scenelet(skel_opt) del skel_opt # skel_dict = skel_opt.to_json() tr_ground[1, 3] = min_y scenelet.aux_info['ground'] = tr_ground.tolist() assert isinstance(ground_rot, list) and len(ground_rot) == 3 scenelet.add_aux_info('ground_rot', ground_rot) scenelet.add_aux_info( 'path_opt_params', { 'rate': args.r, 'w-smooth': args.smooth, 'winsorize-limit': args.winsorize_limit, 'camera': args.camera_name, 'huber': args.huber, 'height_correction': conf.optimize_path.height_correction, 'focal_correction': conf.optimize_path.focal_correction }) scenelet.add_aux_info('frame_ids_filled_in', list(frame_ids_filled_in)) # To MATLAB # _skeleton.get_min_y(_tr_ground) # with skel_opt as skeleton: # skeleton = skel_opt # skeleton_name = os.path.split(args.dir)[0] # skeleton_name = skeleton_name[skeleton_name.rfind('/')+1:] # mdict = skeleton.to_mdict(skeleton_name) # mdict['room_transform'] = tr_ground # mdict['room_transform'][1, 3] *= -1. # print(mdict) # print("scene_name?: %s" % os.path.split(args.dir)[0]) # skeleton.save_matlab( # os.path.join(os.path.dirname(args.dir), "skeleton_opt.mat"), # mdict=mdict) assert scenelet.skeleton.has_forwards(), "No forwards??" scenelet.save(out_path) if show: # save path plot out_path_path = os.path.join(args.dir, os.pardir, "%s_path.jpg" % name_video) path_fig = plot_path(scenelet.skeleton) legend = ["smooth %g" % args.smooth] # hack debug # path_skel2 = os.path.join(args.dir, os.pardir, 'skel_lobby7_nosmooth.json') # if os.path.exists(path_skel2): # skel2 = Skeleton.load(path_skel2) # path_fig = plot_path(skel2, path_fig) # legend.append('no smooth') if show: plt.legend(legend) path_fig.savefig(out_path_path) # backup args path_args = os.path.join(args.dir, os.pardir, 'args_denis.txt') with open(path_args, 'a') as f_args: f_args.write("%s %s\n" % (os.path.basename(sys.executable), " ".join(argv))) # save 2D detections to file if args.postfix == 'unannot': path_skel_ours_2d = os.path.join( args.dir, os.pardir, "skel_%s_2d_%02d.json" % (name_video, 0)) sclt_2d = Scenelet(skel_ours_2d_all) print('Saving {} to {}'.format(len(skel_ours_2d_all.get_frames()), path_skel_ours_2d)) sclt_2d.skeleton.aux_info = {} sclt_2d.save(path_skel_ours_2d) else: print(args.postfix) logging.info("Saving images...") if len(images) and len(out_images): path_out_images = os.path.join(args.dir, os.pardir, 'color') try: os.makedirs(path_out_images) except OSError: pass visible_f_max_log = np.log(visible_f_max) frames = list(out_images.keys()) for frame_id in range(frames[0], frames[-1] + 1): im = out_images[frame_id] if frame_id in out_images \ else cv2.cvtColor(images[frame_id], cv2.COLOR_BGR2RGB) for actor_id in range(args.n_actors): if frame_id in visible_f[actor_id]: frame_id2 = skel_ours_2d_all.unmod_frame_id( frame_id=frame_id, actor_id=actor_id, frames_mod=skel_ours_2d_all.frames_mod) for joint, is_vis in visible_f[actor_id][frame_id].items(): p2d = skel_ours_2d_all.get_joint_3d(joint, frame_id=frame_id2) # radius = np.log(is_vis) / visible_f_max_log # lg.debug("r0: %g" % radius) # radius = np.exp(np.log(is_vis) / visible_f_max_log) # lg.debug("radius is %g" % radius) vis_bool = True if skel_ours_2d_all.has_visible(frame_id=frame_id2, joint_id=joint): vis_bool &= skel_ours_2d_all.is_visible( frame_id2, joint) radius = abs(np.log(is_vis / 0.1 + 1e-6)) if not np.isnan(radius): p2d = (int(round(p2d[0])), int(round(p2d[1]))) cv2.circle(im, center=p2d, radius=int(round(radius)), color=(1., 1., 1., 0.5), thickness=1) conf = get_conf_thresholded(conf=is_vis, thresh_log_conf=None, dtype_np=np.float32) if conf > 0.5: cv2.putText(img=im, text=Joint(joint).get_name(), org=p2d, fontFace=1, fontScale=1, color=(10., 150., 10., 100.)) # lg.debug("set visibility to %g, radius %g" % (is_vis, radius)) # if frame_id in out_images: scale = (shape_orig[1] / float(im.shape[1]), shape_orig[0] / float(im.shape[0])) cv2.imwrite( os.path.join(path_out_images, "color_%05d.jpg" % frame_id), cv2.resize(im, (0, 0), fx=scale[0], fy=scale[1], interpolation=cv2.INTER_CUBIC)) # else: # fname = "color_%05d.jpg" % frame_id # shutil.copyfile( # os.path.join(path_images, fname), # os.path.join(path_out_images, fname)) lg.info("Wrote images to %s/" % path_out_images)