Generate LiDAR Training Dataset

[rogeherlor]

Hi,

I am working with LiDAR DNNs and I want to see the effect of using the LiDAR corruption effect while training instead of using the original ModelNet40 objects. Instead of using your uploaded dataset during the validation stage, I plan to use it during the test stage, and I need to create the LiDAR training set.

When I try to generate the LiDAR training dataset based on the original training set of ModelNet40, ModelNet40-C/data/occlusion.py crashes because core_occlusion() returns a pcd with 0 points:

def core_occlusion(mesh, type, camera_extrinsic=None, camera_intrinsic=None, window_width=1080, window_height=720, n_points=None, downsample_ratio=None):
    if camera_extrinsic is None:
        camera_extrinsic = get_default_camera_extrinsic()
    
    if camera_intrinsic is None:
        camera_intrinsic = get_default_camera_intrinsic()

    camera_parameters = o3d.camera.PinholeCameraParameters()
    camera_parameters.extrinsic = camera_extrinsic
    camera_parameters.intrinsic.set_intrinsics(**camera_intrinsic)

    viewer = o3d.visualization.Visualizer()
    viewer.create_window(width=window_width, height=window_height)
    viewer.add_geometry(mesh)

    control = viewer.get_view_control()
    control.convert_from_pinhole_camera_parameters(camera_parameters)
    # viewer.run()

    depth = viewer.capture_depth_float_buffer(do_render=True)

    viewer.destroy_window()
    pcd = o3d.geometry.PointCloud.create_from_depth_image(depth, camera_parameters.intrinsic, extrinsic=camera_parameters.extrinsic)

    if downsample_ratio is not None:
        ratio =  int((1 - downsample_ratio) / downsample_ratio)
        pcd = pcd.uniform_down_sample(ratio)
    elif n_points is not None:
        # print(np.asarray(pcd.points).shape[0])
        ratio =  int(np.asarray(pcd.points).shape[0] / n_points)
        if ratio > 0:
            # if type == 'occlusion':
            set_points(pcd, shuffle_data(np.asarray(pcd.points)))
            pcd = pcd.uniform_down_sample(ratio)
    
    return pcd


def occlusion_1(mesh, type, severity, window_width=1080, window_height=720, n_points=None, downsample_ratio=None):
    points = get_points(mesh)
    points = normalize(points)
    set_points(mesh, points)
    if type == 'occlusion':
        camera_extrinsic = random_pose(severity)
    elif type == 'lidar':
        camera_extrinsic,pose = lidar_pose(severity)
    camera_intrinsic = get_default_camera_intrinsic(window_width, window_height)
    pcd = core_occlusion(mesh, type, camera_extrinsic=camera_extrinsic, camera_intrinsic=camera_intrinsic, window_width=window_width, window_height=window_height, n_points=n_points, downsample_ratio=downsample_ratio)

    points = get_points(pcd)
    if points.shape[0] < n_points:
        index = np.random.choice(points.shape[0], n_points)
        points = points[index]
    # points = normalize(points)
    # points = denomalize(points, scale, offset)
    if type == 'lidar':
        return points[:n_points,:], pose
    else:
        return points[:n_points,:]

Most of the objects are rendered (i.e.: person_0059.off):

But the ones that return 0 points are not rendered (i.e.: person_0060.off)

Could you upload the training dataset or let me know why some objects return 0 points (i.e.: person_0060.off)? Did this happen when converting the ModelNet40 test split?