PointPillars: Fast Encoders for Object Detection from Point Clouds

A Simple PointPillars PyTorch Implenmentation for 3D Lidar(KITTI) Detection. [Zhihu]

  • It can be run without installing Spconv, mmdet or mmdet3d.
  • Only one detection network (PointPillars) was implemented in this repo, so the code may be more easy to read.
  • Sincere thanks for the great open-souce architectures mmcv, mmdet and mmdet3d, which helps me to learn 3D detetion and implement this repo.

mAP on KITTI validation set (Easy, Moderate, Hard)

Repo Metric Overall Pedestrian Cyclist Car
this repo 3D-BBox 73.3259 62.7834 59.6278 51.4642 47.9446 43.8040 81.8677 63.6617 60.9126 86.6456 76.7439 74.1668
mmdet3d v0.18.1 3D-BBox 72.0537, 60.1114, 55.8320 52.0263, 46.4037, 42.4841 78.7231, 59.9526, 57.2489 85.4118, 73.9780, 67.7630
this repo BEV 77.8540 69.8003 66.6699 59.1687 54.3456 50.5023 84.4268 67.1409 63.7409 89.9664 87.9145 85.7664
mmdet3d v0.18.1 BEV 76.6485, 67.7609, 64.5605 59.0778, 53.3638, 48.4230 80.9328, 63.3447, 60.0618 89.9348, 86.5743, 85.1967
this repo 2D-BBox 80.5097 74.6120 71.4758 64.6249 61.4201 57.5965 86.2569 73.0828 70.1726 90.6471 89.3330 86.6583
mmdet3d v0.18.1 2D-BBox 78.4938, 73.4781, 70.3613 62.2413, 58.9157, 55.3660 82.6460, 72.3547, 68.4669 90.5939, 89.1638, 87.2511
this repo AOS 74.9647 68.1712 65.2817 49.3777 46.7284 43.8352 85.0412 69.1024 66.2801 90.4752 88.6828 85.7298
mmdet3d v0.18.1 AOS 72.41, 66.23, 63.55 46.00, 43.22, 40.94 80.85, 67.20, 63.63 90.37, 88.27, 86.07

Detection Visualization

[Datasets]

  1. Download

    Download point cloud(29GB), images(12 GB), calibration files(16 MB)和labels(5 MB)。Format the datasets as follows:

    kitti
    |- training
        |- calib (#7481 .txt)
        |- image_2 (#7481 .png)
        |- label_2 (#7481 .txt)
        |- velodyne (#7481 .bin)
    |- testing
        |- calib (#7518 .txt)
        |- image_2 (#7518 .png)
        |- velodyne (#7418 .bin)

  2. Pre-process KITTI datasets First

    cd PointPillars/
python pre_process_kitti.py --data_root your_path_to_kitti

    Now, we have datasets as follows:

    kitti
    |- training
        |- calib (#7481 .txt)
        |- image_2 (#7481 .png)
        |- label_2 (#7481 .txt)
        |- velodyne (#7481 .bin)
        |- velodyne_reduced (#7481 .bin)
    |- testing
        |- calib (#7518 .txt)
        |- image_2 (#7518 .png)
        |- velodyne (#7518 .bin)
        |- velodyne_reduced (#7518 .bin)
    |- kitti_gt_database (# 19700 .bin)
    |- kitti_infos_train.pkl
    |- kitti_infos_val.pkl
    |- kitti_infos_trainval.pkl
    |- kitti_infos_test.pkl
    |- kitti_dbinfos_train.pkl

[Compile]

cd ops
python setup.py develop

[Training]

cd PointPillars/
python train.py --data_root your_path_to_kitti

[Evaluation]

cd PointPillars/
python evaluate.py --ckpt pretrained/epoch_160.pth --data_root your_path_to_kitti

[Test]

cd PointPillars/

# 1. infer and visualize point cloud detection
python test.py --ckpt pretrained/epoch_160.pth --pc_path your_pc_path 

# 2. infer and visualize point cloud detection and gound truth.
python test.py --ckpt pretrained/epoch_160.pth --pc_path your_pc_path --calib_path your_calib_path

# 3. infer and visualize point cloud & image detection
python test.py --ckpt pretrained/epoch_160.pth --pc_path your_pc_path --calib_path your_calib_path --img_path your_img_path


e.g. [infer on val set 000134]

python test.py --ckpt pretrained/epoch_160.pth --pc_path /home/lifa/data/KITTI/training/velodyne_reduced/000134.bin

or

python test.py --ckpt pretrained/epoch_160.pth --pc_path /home/lifa/data/KITTI/training/velodyne_reduced/000134.bin --calib_path /home/lifa/data/KITTI/training/calib/000134.txt --img_path /home/lifa/data/KITTI/training/image_2/000134.png --gt_path /home/lifa/data/KITTI/training/label_2/000134.txt

Acknowledements

Thanks for the open souce code mmcv, mmdet and mmdet3d.

GitHub

View Github