Turning pixels into virtual points for multimodal 3D object detection.
Multimodal Virtual Point 3D Detection,
Tianwei Yin, Xingyi Zhou, Philipp Krähenbühl,
arXiv technical report (arXiv 2111.06881 )
@article{yin2021multimodal,
title={Multimodal Virtual Point 3D Detection},
author={Yin, Tianwei and Zhou, Xingyi and Kr{\"a}henb{\"u}hl, Philipp},
journal={NeurIPS},
year={2021},
}
Any questions or suggestions are welcome!
Tianwei Yin yintianwei@utexas.edu Xingyi Zhou zhouxy@cs.utexas.edu
Lidar-based sensing drives current autonomous vehicles. Despite rapid progress, current Lidar sensors still lag two decades behind traditional color cameras in terms of resolution and cost. For autonomous driving, this means that large objects close to the sensors are easily visible, but far-away or small objects comprise only one measurement or two. This is an issue, especially when these objects turn out to be driving hazards. On the other hand, these same objects are clearly visible in onboard RGB sensors. In this work, we present an approach to seamlessly fuse RGB sensors into Lidar-based 3D recognition. Our approach takes a set of 2D detections to generate dense 3D virtual points to augment an otherwise sparse 3D point-cloud. These virtual points naturally integrate into any standard Lidar-based 3D detectors along with regular Lidar measurements. The resulting multi-modal detector is simple and effective. Experimental results on the large-scale nuScenes dataset show that our framework improves a strong CenterPoint baseline by a significant 6.6 mAP, and outperforms competing fusion approaches.
| MAP ↑ | NDS ↑ | |
|---|---|---|
| CenterPoint-Voxel | 59.5 | 66.7 |
| CenterPoint-Voxel + MVP | 66.0 | 69.9 |
| CenterPoint-Pillar | 52.4 | 61.5 |
| CenterPoint-Pillar + MVP | 62.8 | 66.2 |
| MAP ↑ | NDS ↑ | PKL ↓ | |
|---|---|---|---|
| MVP | 66.4 | 70.5 | 0.603 |
Please install CenterPoint and CenterNet2. Make sure to add a link to CenterNet2 folder in your python path. We will use CenterNet2 for 2D instance segmentation and CenterPoint for 3D detection.
# For nuScenes Dataset
└── NUSCENES_DATASET_ROOT
├── samples <-- key frames
├── sweeps <-- frames without annotation
├── maps <-- unused
├── v1.0-trainval <-- metadata
Create a symlink to the dataset root in both CenterPoint and MVP's root directories.
mkdir data && cd data
ln -s DATA_ROOT nuScenesRemember to change the DATA_ROOT to the actual path in your system.
You can download the precomputed virtual points from here.
If you prefer to generating the virtual points yourself. Download the centernet2 model from here and place it in the root directory.
Use the following command in the current directory to generate virtual points for nuscenes training and validation sets. The points will be saved to data/nuScenes/samples or sweeps/LIDAR_TOP_VIRTUAL.
python virtual_gen.py --info_path data/nuScenes/infos_train_10sweeps_withvelo_filter_True.pkl MODEL.WEIGHTS centernet2_checkpoint.pth You will need about 80GB space and the whole process will take 10 to 20 hours using a single GPU.
Go to the CenterPoint's root directory and run
# nuScenes
python tools/create_data.py nuscenes_data_prep --root_path=NUSCENES_TRAINVAL_DATASET_ROOT --version="v1.0-trainval" --nsweeps=10 --virtual True
To save time and space, before running the above command, you can also remove lines below https://github.com/tianweiy/CenterPoint/blob/master/tools/create_data.py#L13 to avoid generating gt database. After that, remember to set https://github.com/tianweiy/CenterPoint/blob/master/configs/mvp/nusc_centerpoint_pp_fix_bn_z_scale_virtual.py#L135 to None. The improvements of gt sampling on nuscenes is marginal (<0.5nds).
if you want to reproduce CenterPoint baseline's results, then also run the following command
# nuScenes
python tools/create_data.py nuscenes_data_prep --root_path=NUSCENES_TRAINVAL_DATASET_ROOT --version="v1.0-trainval" --nsweeps=10 --virtual False
In the end, the data and info files should be organized as follows
# For nuScenes Dataset
└── CenterPoint
└── data
└── nuScenes
├── maps <-- unused
|── v1.0-trainval <-- metadata and annotations
|── infos_train_10sweeps_withvelo_filter_True.pkl <-- train annotations
|── infos_val_10sweeps_withvelo_filter_True.pkl <-- val annotations
|── dbinfos_train_10sweeps_withvelo_virtual.pkl <-- GT database info files
|── gt_database_10sweeps_withvelo_virtual <-- GT database
|── samples <-- key frames
|── LIDAR_TOP
|── LIDAR_TOP_VIRTUAL
└── sweeps <-- frames without annotation
|── LIDAR_TOP
|── LIDAR_TOP_VIRTUAL
Go to CenterPoint's root directory and use the following command to start a distributed training using 4 GPUs. The models and logs will be saved to work_dirs/CONFIG_NAME
python -m torch.distributed.launch --nproc_per_node=4 ./tools/train.py CONFIG_PATHFor distributed testing with 4 gpus,
python -m torch.distributed.launch --nproc_per_node=4 ./tools/dist_test.py CONFIG_PATH --work_dir work_dirs/CONFIG_NAME --checkpoint work_dirs/CONFIG_NAME/latest.pth For testing with one gpu and see the inference time,
python ./tools/dist_test.py CONFIG_PATH --work_dir work_dirs/CONFIG_NAME --checkpoint work_dirs/CONFIG_NAME/latest.pth --speed_test We experiment with VoxelNet and PointPillars architectures on nuScenes.
| Model | Validation MAP | Validation NDS | Link |
|---|---|---|---|
| centerpoint_baseline | 59.5 | 66.7 | URL |
| Ours | 66.0 | 69.9 | URL |
| Ours + Two Stage | 67.0 | 70.7 | URL |
| Model | Validation MAP | Validation NDS | Link |
|---|---|---|---|
| centerpoint_baseline | 52.4 | 61.5 | URL |
| Ours | 62.8 | 66.2 | URL |
MIT License.