Home

Awesome

SparseOcc

This is the official PyTorch implementation for our paper:

Fully Sparse 3D Panoptic Occupancy Prediction<br> :school: Presented by Nanjing University and Shanghai AI Lab<br> :email: Primary contact: Haisong Liu (afterthat97@gmail.com)<br> :trophy: CVPR 2024 Autonomous Driving Challenge - Occupancy and Flow<br> :book: 中文解读（官方）：https://zhuanlan.zhihu.com/p/709576252<br> :book: 中文解读（第三方）: AIming, 自动驾驶之心

:warning: Important Notes

There is another concurrent project titled ''SparseOcc: Rethinking sparse latent representation for vision-based semantic occupancy prediction'' by Tang et al., which shares the same name SparseOcc with ours. However, this repository is unrelated to the aforementioned paper.

If you cite our research, please ensure that you reference the correct version (arXiv 2312.17118, authored by Liu et al.):

@article{liu2023fully,
  title={Fully sparse 3d panoptic occupancy prediction},
  author={Liu, Haisong and Wang, Haiguang and Chen, Yang and Yang, Zetong and Zeng, Jia and Chen, Li and Wang, Limin},
  journal={arXiv preprint arXiv:2312.17118},
  year={2023}
}

In arXiv 2312.17118v3, we removed the word "panoptic" from the title. However, Google Scholar's database has not been updated and still shows the old one. Therefore, we still recommend citing the old title - "Fully sparse 3d panoptic occupancy prediction" - so that Google Scholar can index it correctly. Thank you all.

News

Highlights

New model:1st_place_medal:: SparseOcc initially reconstructs a sparse 3D representation from visual inputs and subsequently predicts semantic/instance occupancy from the 3D sparse representation by sparse queries.

New evaluation metric:chart_with_upwards_trend:: We design a thoughtful ray-based evaluation metric, namely RayIoU, to solve the inconsistency penalty along depths raised in traditional voxel-level mIoU criteria.

Some FAQs from the community about the evaluation metrics:

  1. Why does training with visible masks result in significant improvements in the old mIoU metric, but not in the new RayIoU metric? As mentioned in the paper, when using the visible mask during training, the area behind the surface won't be supervised, so the model tends to fill this area with duplicated predictions, leading to a thicker surface. The old metric inconsistently penalizes along the depth axis when the prediction has a thick surface. Thus, this ''imporovement'' is mainly due to the vulnerability of old metric.
  2. Why SparseOcc cannot exploit the vulnerability of the old metrics? As SparseOcc employs a fully sparse architecture, it always predicts a thin surface. Thus, there are two ways for a fair comparison: (a) use the old metric, but all methods must predict a thin surface, which implies they cannot use the visible mask during training; (b) use RayIoU, as it is more reasonable and can fairly compare thick or thin surface. Our method achieves SOTA performance on both cases.
  3. Does RayIoU overlook interior reconstruction? Firstly, we are unable to obtain the interior occupancy ground-truth. This is because the ground-truth is derived from voxelizing LiDAR point clouds, and LiDARs are only capable of scanning the thin surface of an object. Secondly, the query ray in RayIoU can originate from any position within the scene (see the figure above). This allows it to evaluate the overall reconstruction performance, unlike depth estimation. We would like to emphasize that the evaluation logic of RayIoU aligns with the process of ground-truth generation.

If you have other questions, feel free to contact me (Haisong Liu, afterthat97@gmail.com).

Model Zoo

These results are from our latest version, v1.1, which outperforms the results reported in the paper. Additionally, our implementation differs slightly from the original paper. If you wish to reproduce the paper exactly, please refer to the v1.0 tag.

SettingEpochsTraining CostRayIoURayPQFPSWeights
r50_nuimg_704x256_8f2415h, ~12GB36.8-17.3github
r50_nuimg_704x256_8f_60e6037h, ~12GB37.7-17.3github
r50_nuimg_704x256_8f_pano2415h, ~12GB35.914.017.3github

Environment

The requirements are the same as those of SparseBEV.

Install PyTorch 2.0 + CUDA 11.8:

conda create -n sparseocc python=3.8
conda activate sparseocc
conda install pytorch==2.0.0 torchvision==0.15.0 pytorch-cuda=11.8 -c pytorch -c nvidia

Install other dependencies:

pip install openmim
mim install mmcv-full==1.6.0
mim install mmdet==2.28.2
mim install mmsegmentation==0.30.0
mim install mmdet3d==1.0.0rc6
pip install setuptools==59.5.0
pip install numpy==1.23.5

Install turbojpeg and pillow-simd to speed up data loading (optional but important):

sudo apt-get update
sudo apt-get install -y libturbojpeg
pip install pyturbojpeg
pip uninstall pillow
pip install pillow-simd==9.0.0.post1

Compile CUDA extensions:

cd models/csrc
python setup.py build_ext --inplace

Prepare Dataset

The first two steps are the same as those of SparseBEV.

  1. Download nuScenes from https://www.nuscenes.org/nuscenes, put it to data/nuscenes and preprocess it with mmdetection3d.

  2. Download the generated info file from gdrive and unzip it. These *.pkl files can also be generated with our script: gen_sweep_info.py.

  3. Download Occ3D-nuScenes occupancy GT from gdrive, unzip it, and save it to data/nuscenes/occ3d.

  4. Folder structure:

data/nuscenes
├── maps
├── nuscenes_infos_test_sweep.pkl
├── nuscenes_infos_train_sweep.pkl
├── nuscenes_infos_val_sweep.pkl
├── samples
├── sweeps
├── v1.0-test
└── v1.0-trainval
└── occ3d
    ├── scene-0001
    │   ├── 0037a705a2e04559b1bba6c01beca1cf
    │   │   └── labels.npz
    │   ├── 026155aa1c554e2f87914ec9ba80acae
    │   │   └── labels.npz
    ...
  1. (Optional) Generate the panoptic occupancy ground truth with gen_instance_info.py. The panoptic version of Occ3D will be saved to data/nuscenes/occ3d_panoptic.

Training

Train SparseOcc with 8 GPUs:

torchrun --nproc_per_node 8 train.py --config configs/sparseocc_r50_nuimg_704x256_8f.py

Train SparseOcc with 4 GPUs (i.e the last four GPUs):

export CUDA_VISIBLE_DEVICES=4,5,6,7
torchrun --nproc_per_node 4 train.py --config configs/sparseocc_r50_nuimg_704x256_8f.py

The batch size for each GPU will be scaled automatically. So there is no need to modify the batch_size in config files.

Evaluation

Single-GPU evaluation:

export CUDA_VISIBLE_DEVICES=0
python val.py --config configs/sparseocc_r50_nuimg_704x256_8f.py --weights checkpoints/sparseocc_r50_nuimg_704x256_8f.pth

Multi-GPU evaluation:

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
torchrun --nproc_per_node 8 val.py --config configs/sparseocc_r50_nuimg_704x256_8f.py --weights checkpoints/sparseocc_r50_nuimg_704x256_8f.pth

Standalone Evaluation

If you want to evaluate your own model using RayIoU, please follow the steps below:

  1. Save the predictions (shape=[200x200x16], dtype=np.uint8) with the compressed npz format. For example:
save_path = os.path.join(save_dir, sample_token + '.npz')
np.savez_compressed(save_path, pred=sem_pred)
  1. The filename for each sample is sample_token.npz, for example:
prediction/your_model
├── 000681a060c04755a1537cf83b53ba57.npz
├── 000868a72138448191b4092f75ed7776.npz
├── 0017c2623c914571a1ff2a37f034ffd7.npz
├── ...
  1. Run ray_metrics.py to evaluate on the RayIoU:
python ray_metrics.py --pred-dir prediction/your_model

Timing

FPS is measured with a single GPU:

export CUDA_VISIBLE_DEVICES=0
python timing.py --config configs/sparseocc_r50_nuimg_704x256_8f.py --weights checkpoints/sparseocc_r50_nuimg_704x256_8f.pth

Acknowledgements

Many thanks to these excellent open-source projects: