Custom Dataset Tutorial

For the custom dataset template, we only consider the basic scenario: raw point clouds and their corresponding annotations. Point clouds are supposed to be stored in .npy format.

Label format

We only consider the most basic information -- category and bounding box in the label template. Annotations are stored in the .txt. Each line represents a box in a given scene as below:

# format: [x y z dx dy dz heading_angle category_name]
1.50 1.46 0.10 5.12 1.85 4.13 1.56 Vehicle
5.54 0.57 0.41 1.08 0.74 1.95 1.57 Pedestrian

The box should in the unified 3D box definition (see README)

Files structure

Files should be placed as the following folder structure:

OpenPCDet
├── data
│   ├── custom
│   │   │── ImageSets
│   │   │   │── train.txt
│   │   │   │── val.txt
│   │   │── points
│   │   │   │── 000000.npy
│   │   │   │── 999999.npy
│   │   │── labels
│   │   │   │── 000000.txt
│   │   │   │── 999999.txt
├── pcdet
├── tools

Dataset splits need to be pre-defined and placed in ImageSets

Hyper-parameters Configurations

Point cloud features

Modify following configurations in custom_dataset.yaml to suit your own point clouds.

POINT_FEATURE_ENCODING: {
    encoding_type: absolute_coordinates_encoding,
    used_feature_list: ['x', 'y', 'z', 'intensity'],
    src_feature_list: ['x', 'y', 'z', 'intensity'],
}
...
# In gt_sampling data augmentation
NUM_POINT_FEATURES: 4

Point cloud range and voxel sizes

For voxel based detectors such as SECOND, PV-RCNN and CenterPoint, the point cloud range and voxel size should follow:

Point cloud range along z-axis / voxel_size is 40
Point cloud range along x&y-axis / voxel_size is the multiple of 16.

Notice that the second rule also suit pillar based detectors such as PointPillar and CenterPoint-Pillar.

Category names and anchor sizes

Category names and anchor size are need to be adapted to custom datasets.

CLASS_NAMES: ['Vehicle', 'Pedestrian', 'Cyclist']  
...
MAP_CLASS_TO_KITTI: {
   'Vehicle': 'Car',
   'Pedestrian': 'Pedestrian',
   'Cyclist': 'Cyclist',
}
...
'anchor_sizes': [[3.9, 1.6, 1.56]],
...
# In gt sampling data augmentation
PREPARE: {
filter_by_min_points: ['Vehicle:5', 'Pedestrian:5', 'Cyclist:5'],
filter_by_difficulty: [-1],
}
SAMPLE_GROUPS: ['Vehicle:20','Pedestrian:15', 'Cyclist:15']
...

In addition, please also modify the default category names for creating infos in custom_dataset.py

create_custom_infos(
    dataset_cfg=dataset_cfg,
    class_names=['Vehicle', 'Pedestrian', 'Cyclist'],
    data_path=ROOT_DIR / 'data' / 'custom',
    save_path=ROOT_DIR / 'data' / 'custom',
)

Create data info

Generate the data infos by running the following command:

python -m pcdet.datasets.custom.custom_dataset create_custom_infos tools/cfgs/dataset_configs/custom_dataset.yaml

Evaluation

Here, we only provide an implementation for KITTI stype evaluation. The category mapping between custom dataset and KITTI need to be defined in the custom_dataset.yaml

MAP_CLASS_TO_KITTI: {
    'Vehicle': 'Car',
    'Pedestrian': 'Pedestrian',
    'Cyclist': 'Cyclist',
}

3.3 KiB Raw Blame History