def forward(self, obs_local, valid_points): # obs_local: <BxHxWx3> # valid_points: <BxHxW> self.obs_local = deepcopy(obs_local) self.valid_points = valid_points self.pose_est = self.loc_net(self.obs_local) bs = obs_local.shape[0] self.obs_local = self.obs_local.view(bs, -1, 3) self.valid_points = self.valid_points.view(bs, -1) self.obs_global_est = transform_to_global_AVD(self.pose_est, self.obs_local) if self.training: self.unoccupied_local = sample_unoccupied_point( self.obs_local, self.n_samples) self.unoccupied_global = transform_to_global_AVD( self.pose_est, self.unoccupied_local) inputs, self.gt = get_M_net_inputs_labels(self.obs_global_est, self.unoccupied_global) self.occp_prob = self.occup_net(inputs) loss = self.compute_loss() return loss
def __getitem__(self, index): pcd = self.point_clouds[index, :, :, :] # <HxWx3> valid_points = self.valid_points[index, :] #<HxW> if self._trans_by_pose is not None: pcd = pcd.unsqueeze(0) # <1XHxWx3> pose = self._trans_by_pose[index, :].unsqueeze(0) # <1x3> pcd = utils.transform_to_global_AVD(pose, pcd).squeeze(0) return pcd, valid_points
def get_scores(checkpoint_dir): saved_json_file = os.path.join(checkpoint_dir, 'opt.json') train_opt = utils.load_opt_from_json(saved_json_file) name = train_opt['name'] data_dir = '../../../data/ActiveVisionDataset/' + train_opt[ 'data_dir'].split('/')[-1] subsample_rate = train_opt['subsample_rate'] traj = train_opt['traj'] print(data_dir) # load ground truth poses dataset = AVD(data_dir, traj, subsample_rate) gt_pose = dataset.gt gt_location = gt_pose[:, :3] pts = dataset.point_clouds #.numpy() #np.save("local.npy",pts) #np.save("pose.npy",gt_pose) # load predicted poses pred_file = os.path.join(opt.checkpoint_dir, 'pose_est.npy') pred_pose = np.load(pred_file) pred_location = pred_pose[:, :2] * dataset.depth_scale # denormalization pred_location = add_y_coord_for_evaluation(pred_location) #print(gt_pose) print(pred_pose) ate, aligned_location = utils.compute_ate(pred_location, gt_location) print('{}, ate: {}'.format(name, ate)) gt_pose[:, :3] = gt_pose[:, :3] / dataset.depth_scale gt_yaw = np.arccos( gt_pose[:, 5] / np.sqrt(gt_pose[:, 3] * gt_pose[:, 3] + gt_pose[:, 5] * gt_pose[:, 5])) gt_pose_xzth = np.vstack((gt_pose[:, 0], gt_pose[:, 2], -gt_yaw)).transpose() colors = np.array([[0, 1, 1], [0, 0, 0], [0, 0, 1], [1, 0, 1], [0.5, 0.5, 0.5], [0, 0.5, 0], [0, 1, 0], [0.5, 0, 0], [0, 0, 0.5], [0.5, 0.5, 0], [0.5, 0, 0.5], [1, 0, 0], [0.75, 0.75, 0.75], [0, 0.5, 0.5], [1, 1, 1], [1, 1, 0]]) #path_or = '../../../bk_origin/DeepMapping/' global_point_cloud_file = os.path.join(opt.checkpoint_dir, 'obs_global_est.npy') #global_point_cloud_file_or = os.path.join(path_or,global_point_cloud_file[3:]) pcds_ours = utils.load_obs_global_est(global_point_cloud_file, colors) #pcds_or = utils.load_obs_global_est(global_point_cloud_file_or,colors) pts_gt = transform_to_global_AVD(gt_pose_xzth, pts).numpy() #pts_or = np.load(global_point_cloud_file_or) pts_ours = np.load(global_point_cloud_file) pts_gt = pts_gt.reshape((16, -1, 3)) print(pt_diff(pts_gt, pts_ours)) print(pt_diff(pts_gt, pts_or))