ScanNet for 3D Object Detection

Dataset preparation

For the overall process, please refer to the README page for ScanNet.

Export ScanNet point cloud data

By exporting ScanNet data, we load the raw point cloud data and generate the relevant annotations including semantic labels, instance labels and ground truth bounding boxes.


The directory structure before data preparation should be as below

├── mmdet3d
├── tools
├── configs
├── data
│   ├── scannet
│   │   ├── meta_data
│   │   ├── scans
│   │   │   ├── scenexxxx_xx
│   │   ├──
│   │   ├──
│   │   ├──
│   │   ├──

Under folder scans there are overall 1201 train and 312 validation folders in which raw point cloud data and relevant annotations are saved. For instance, under folder scene0001_01 the files are as below:

  • scene0001_01_vh_clean_2.ply: Mesh file storing coordinates and colors of each vertex. The mesh’s vertices are taken as raw point cloud data.

  • scene0001_01.aggregation.json: Aggregation file including object ID, segments ID and label.

  • scene0001_01_vh_clean_2.0.010000.segs.json: Segmentation file including segments ID and vertex.

  • scene0001_01.txt: Meta file including axis-aligned matrix, etc.

  • scene0001_01_vh_clean_2.labels.ply: Annotation file containing the category of each vertex.

Export ScanNet data by running python The main steps include:

  • Export original files to point cloud, instance label, semantic label and bounding box file.

  • Downsample raw point cloud and filter invalid classes.

  • Save point cloud data and relevant annotation files.

And the core function export in is as follows:

def export(mesh_file,

    # label map file: ./data/scannet/meta_data/scannetv2-labels.combined.tsv
    # the various label standards in the label map file, e.g. 'nyu40id'
    label_map = scannet_utils.read_label_mapping(
        label_map_file, label_from='raw_category', label_to='nyu40id')
    # load raw point cloud data, 6-dims feature: XYZRGB
    mesh_vertices = scannet_utils.read_mesh_vertices_rgb(mesh_file)

    # Load scene axis alignment matrix: a 4x4 transformation matrix
    # transform raw points in sensor coordinate system to a coordinate system
    # which is axis-aligned with the length/width of the room
    lines = open(meta_file).readlines()
    # test set data doesn't have align_matrix
    axis_align_matrix = np.eye(4)
    for line in lines:
        if 'axisAlignment' in line:
            axis_align_matrix = [
                for x in line.rstrip().strip('axisAlignment = ').split(' ')
    axis_align_matrix = np.array(axis_align_matrix).reshape((4, 4))

    # perform global alignment of mesh vertices
    pts = np.ones((mesh_vertices.shape[0], 4))
    # raw point cloud in homogeneous coordinates, each row: [x, y, z, 1]
    pts[:, 0:3] = mesh_vertices[:, 0:3]
    # transform raw mesh vertices to aligned mesh vertices
    pts =, axis_align_matrix.transpose())  # Nx4
    aligned_mesh_vertices = np.concatenate([pts[:, 0:3], mesh_vertices[:, 3:]],

    # Load semantic and instance labels
    if not test_mode:
        # each object has one semantic label and consists of several segments
        object_id_to_segs, label_to_segs = read_aggregation(agg_file)
        # many points may belong to the same segment
        seg_to_verts, num_verts = read_segmentation(seg_file)
        label_ids = np.zeros(shape=(num_verts), dtype=np.uint32)
        object_id_to_label_id = {}
        for label, segs in label_to_segs.items():
            label_id = label_map[label]
            for seg in segs:
                verts = seg_to_verts[seg]
                # each point has one semantic label
                label_ids[verts] = label_id
        instance_ids = np.zeros(
            shape=(num_verts), dtype=np.uint32)  # 0: unannotated
        for object_id, segs in object_id_to_segs.items():
            for seg in segs:
                verts = seg_to_verts[seg]
                # object_id is 1-indexed, i.e. 1,2,3,.,,,.NUM_INSTANCES
                # each point belongs to one object
                instance_ids[verts] = object_id
                if object_id not in object_id_to_label_id:
                    object_id_to_label_id[object_id] = label_ids[verts][0]
        # bbox format is [x, y, z, dx, dy, dz, label_id]
        # [x, y, z] is gravity center of bbox, [dx, dy, dz] is axis-aligned
        # [label_id] is semantic label id in 'nyu40id' standard
        # Note: since 3D bbox is axis-aligned, the yaw is 0.
        unaligned_bboxes = extract_bbox(mesh_vertices, object_id_to_segs,
                                        object_id_to_label_id, instance_ids)
        aligned_bboxes = extract_bbox(aligned_mesh_vertices, object_id_to_segs,
                                      object_id_to_label_id, instance_ids)

    return mesh_vertices, label_ids, instance_ids, unaligned_bboxes, \
        aligned_bboxes, object_id_to_label_id, axis_align_matrix

After exporting each scan, the raw point cloud could be downsampled, e.g. to 50000, if the number of points is too large (the raw point cloud won’t be downsampled if it’s also used in 3D semantic segmentation task). In addition, invalid semantic labels outside of nyu40id standard or optional DONOT CARE classes should be filtered. Finally, the point cloud data, semantic labels, instance labels and ground truth bounding boxes should be saved in .npy files.

Export ScanNet RGB data (optional)

By exporting ScanNet RGB data, for each scene we load a set of RGB images with corresponding 4x4 pose matrices, and a single 4x4 camera intrinsic matrix. Note, that this step is optional and can be skipped if multi-view detection is not planned to use.


Each of 1201 train, 312 validation and 100 test scenes contains a single .sens file. For instance, for scene 0001_01 we have data/scannet/scans/scene0001_01/0001_01.sens. For this scene all images and poses are extracted to data/scannet/posed_images/scene0001_01. Specifically, there will be 300 image files xxxxx.jpg, 300 camera pose files xxxxx.txt and a single intrinsic.txt file. Typically, single scene contains several thousand images. By default, we extract only 300 of them with resulting space occupation of <100 Gb. To extract more images, use --max-images-per-scene parameter.

Create dataset

python tools/ scannet --root-path ./data/scannet \
--out-dir ./data/scannet --extra-tag scannet

The above exported point cloud file, semantic label file and instance label file are further saved in .bin format. Meanwhile .pkl info files are also generated for train or validation. The core function process_single_scene of getting data infos is as follows.

def process_single_scene(sample_idx):

    # save point cloud, instance label and semantic label in .bin file respectively, get info['pts_path'], info['pts_instance_mask_path'] and info['pts_semantic_mask_path']

    # get annotations
    if has_label:
        annotations = {}
        # box is of shape [k, 6 + class]
        aligned_box_label = self.get_aligned_box_label(sample_idx)
        unaligned_box_label = self.get_unaligned_box_label(sample_idx)
        annotations['gt_num'] = aligned_box_label.shape[0]
        if annotations['gt_num'] != 0:
            aligned_box = aligned_box_label[:, :-1]  # k, 6
            unaligned_box = unaligned_box_label[:, :-1]
            classes = aligned_box_label[:, -1]  # k
            annotations['name'] = np.array([
                for i in range(annotations['gt_num'])
            # default names are given to aligned bbox for compatibility
            # we also save unaligned bbox info with marked names
            annotations['location'] = aligned_box[:, :3]
            annotations['dimensions'] = aligned_box[:, 3:6]
            annotations['gt_boxes_upright_depth'] = aligned_box
            annotations['unaligned_location'] = unaligned_box[:, :3]
            annotations['unaligned_dimensions'] = unaligned_box[:, 3:6]
                'unaligned_gt_boxes_upright_depth'] = unaligned_box
            annotations['index'] = np.arange(
                annotations['gt_num'], dtype=np.int32)
            annotations['class'] = np.array([
                for i in range(annotations['gt_num'])
        axis_align_matrix = self.get_axis_align_matrix(sample_idx)
        annotations['axis_align_matrix'] = axis_align_matrix  # 4x4
        info['annos'] = annotations
    return info

The directory structure after process should be as below

├── meta_data
├── scans
├── scans_test
├── scannet_instance_data
├── points
│   ├── xxxxx.bin
├── instance_mask
│   ├── xxxxx.bin
├── semantic_mask
│   ├── xxxxx.bin
├── seg_info
│   ├── train_label_weight.npy
│   ├── train_resampled_scene_idxs.npy
│   ├── val_label_weight.npy
│   ├── val_resampled_scene_idxs.npy
├── posed_images
│   ├── scenexxxx_xx
│   │   ├── xxxxxx.txt
│   │   ├── xxxxxx.jpg
│   │   ├── intrinsic.txt
├── scannet_infos_train.pkl
├── scannet_infos_val.pkl
├── scannet_infos_test.pkl
  • points/xxxxx.bin: The axis-unaligned point cloud data after downsample. Since ScanNet 3D detection task takes axis-aligned point clouds as input, while ScanNet 3D semantic segmentation task takes unaligned points, we choose to store unaligned points and their axis-align transform matrix. Note: the points would be axis-aligned in pre-processing pipeline GlobalAlignment of 3D detection task.

  • instance_mask/xxxxx.bin: The instance label for each point, value range: [0, NUM_INSTANCES], 0: unannotated.

  • semantic_mask/xxxxx.bin: The semantic label for each point, value range: [1, 40], i.e. nyu40id standard. Note: the nyu40id ID will be mapped to train ID in train pipeline PointSegClassMapping.

  • posed_images/scenexxxx_xx: The set of .jpg images with .txt 4x4 poses and the single .txt file with camera intrinsic matrix.

  • scannet_infos_train.pkl: The train data infos, the detailed info of each scan is as follows:

    • info[‘point_cloud’]: {‘num_features’: 6, ‘lidar_idx’: sample_idx}.

    • info[‘pts_path’]: The path of points/xxxxx.bin.

    • info[‘pts_instance_mask_path’]: The path of instance_mask/xxxxx.bin.

    • info[‘pts_semantic_mask_path’]: The path of semantic_mask/xxxxx.bin.

    • info[‘annos’]: The annotations of each scan.

      • annotations[‘gt_num’]: The number of ground truths.

      • annotations[‘name’]: The semantic name of all ground truths, e.g. chair.

      • annotations[‘location’]: The gravity center of the axis-aligned 3D bounding boxes in depth coordinate system. Shape: [K, 3], K is the number of ground truths.

      • annotations[‘dimensions’]: The dimensions of the axis-aligned 3D bounding boxes in depth coordinate system, i.e. (x_size, y_size, z_size), shape: [K, 3].

      • annotations[‘gt_boxes_upright_depth’]: The axis-aligned 3D bounding boxes in depth coordinate system, each bounding box is (x, y, z, x_size, y_size, z_size), shape: [K, 6].

      • annotations[‘unaligned_location’]: The gravity center of the axis-unaligned 3D bounding boxes in depth coordinate system.

      • annotations[‘unaligned_dimensions’]: The dimensions of the axis-unaligned 3D bounding boxes in depth coordinate system.

      • annotations[‘unaligned_gt_boxes_upright_depth’]: The axis-unaligned 3D bounding boxes in depth coordinate system.

      • annotations[‘index’]: The index of all ground truths, i.e. [0, K).

      • annotations[‘class’]: The train class ID of the bounding boxes, value range: [0, 18), shape: [K, ].

  • scannet_infos_val.pkl: The val data infos, which shares the same format as scannet_infos_train.pkl.

  • scannet_infos_test.pkl: The test data infos, which almost shares the same format as scannet_infos_train.pkl except for the lack of annotation.

Training pipeline

A typical training pipeline of ScanNet for 3D detection is as follows.

train_pipeline = [
        use_dim=[0, 1, 2]),
    dict(type='GlobalAlignment', rotation_axis=2),
        valid_cat_ids=(3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 16, 24, 28, 33, 34,
                       36, 39),
    dict(type='PointSample', num_points=40000),
        rot_range=[-0.087266, 0.087266],
        scale_ratio_range=[1.0, 1.0],
    dict(type='DefaultFormatBundle3D', class_names=class_names),
            'points', 'gt_bboxes_3d', 'gt_labels_3d', 'pts_semantic_mask',
  • GlobalAlignment: The previous point cloud would be axis-aligned using the axis-aligned matrix.

  • PointSegClassMapping: Only the valid category IDs will be mapped to class label IDs like [0, 18) during training.

  • Data augmentation:

    • PointSample: downsample the input point cloud.

    • RandomFlip3D: randomly flip the input point cloud horizontally or vertically.

    • GlobalRotScaleTrans: rotate the input point cloud, usually in the range of [-5, 5] (degrees) for ScanNet; then scale the input point cloud, usually by 1.0 for ScanNet (which means no scaling); finally translate the input point cloud, usually by 0 for ScanNet (which means no translation).


Typically mean Average Precision (mAP) is used for evaluation on ScanNet, e.g. mAP@0.25 and mAP@0.5. In detail, a generic function to compute precision and recall for 3D object detection for multiple classes is called, please refer to indoor_eval.

As introduced in section Export ScanNet data, all ground truth 3D bounding box are axis-aligned, i.e. the yaw is zero. So the yaw target of network predicted 3D bounding box is also zero and axis-aligned 3D Non-Maximum Suppression (NMS), which is regardless of rotation, is adopted during post-processing .