fix(nuscenes): ensure self.scene is set when NuScenes (+3 more)

[andrewwhitecdw]

Small fixes in datasets/nuscenes.py:

fix: ensure self.scene is set when NuScenes handle is passed in

Fix: Replace:

        if self.nusc is None:
            self.nusc = NuScenes(
                version="v1.0-trainval", dataroot=self.data_path, verbose=True
            )
            self.scene = self.nusc.scene[self.scene_idx]
        total_camera_list = [

with:

        if self.nusc is None:
            self.nusc = NuScenes(
                version="v1.0-trainval", dataroot=self.data_path, verbose=True
            )
        self.scene = self.nusc.scene[self.scene_idx]
        total_camera_list = [

fix: truncate lidar rays by Euclidean range, not world x coordinate

Fix: Replace:

            valid_mask = torch.ones_like(lidar_origins[:, 0]).bool()
            if self.data_cfg.truncated_max_range is not None:
                valid_mask = lidar_points[:, 0] < self.data_cfg.truncated_max_range
            if self.data_cfg.truncated_min_range is not None:
                valid_mask = valid_mask & (
                    lidar_points[:, 0] > self.data_cfg.truncated_min_range
                )

with:

            valid_mask = torch.ones_like(lidar_origins[:, 0]).bool()
            if self.data_cfg.truncated_max_range is not None:
                valid_mask = lidar_ranges.squeeze(-1) < self.data_cfg.truncated_max_range
            if self.data_cfg.truncated_min_range is not None:
                valid_mask = valid_mask & (
                    lidar_ranges.squeeze(-1) > self.data_cfg.truncated_min_range
                )

fix: clamp normalized feature colors to [0,1] after division

Fix: Replace:

                features = (features - self.pixel_source.feat_color_min) / (
                    self.pixel_source.feat_color_max - self.pixel_source.feat_color_min
                ).clamp(0, 1)

with:

                features = ((features - self.pixel_source.feat_color_min) / (
                    self.pixel_source.feat_color_max - self.pixel_source.feat_color_min
                )).clamp(0, 1)

fix: count original lidar rays before remove_close filtering

Fix: Replace:

            lidar_pc = LidarPointCloud.from_file(self.lidar_filepaths[t])
            lidar_pc.remove_close(1.0)
            pc = lidar_pc.points[:3, :].T
            pc = np.hstack((pc, np.ones((pc.shape[0], 1))))
            pc = torch.from_numpy(pc).float()
            lidar_points = pc @ self.lidar_to_worlds[t].T
            lidar_points = lidar_points[:, :3]
            lidar_origins = (
                self.lidar_to_worlds[t][:3, 3]
                .unsqueeze(0)
                .repeat(lidar_points.shape[0], 1)
            )
            lidar_directions = lidar_points - lidar_origins
            lidar_ranges = torch.norm(lidar_directions, dim=-1, keepdim=True)
            lidar_directions = lidar_directions / lidar_ranges
            accumulated_num_original_rays += len(lidar_pc.points[0])

with:

            lidar_pc = LidarPointCloud.from_file(self.lidar_filepaths[t])
            accumulated_num_original_rays += len(lidar_pc.points[0])
            lidar_pc.remove_close(1.0)
            pc = lidar_pc.points[:3, :].T
            pc = np.hstack((pc, np.ones((pc.shape[0], 1))))
            pc = torch.from_numpy(pc).float()
            lidar_points = pc @ self.lidar_to_worlds[t].T
            lidar_points = lidar_points[:, :3]
            lidar_origins = (
                self.lidar_to_worlds[t][:3, 3]
                .unsqueeze(0)
                .repeat(lidar_points.shape[0], 1)
            )
            lidar_directions = lidar_points - lidar_origins
            lidar_ranges = torch.norm(lidar_directions, dim=-1, keepdim=True)
            lidar_directions = lidar_directions / lidar_ranges

Files changed

• datasets/nuscenes.py

[andrewwhitecdw]

Closing this sweep-generated PR: sole commit is missing a valid Signed-off-by trailer. It does not meet the sweep requirements (single signed-off commit).