TPVFormer: An academic alternative to Tesla's Occupancy Network

Paper | Project Page | Leaderboard

Tri-Perspective View for Vision-Based 3D Semantic Occupancy Prediction, CVPR 2023

Yuanhui Huang*, Wenzhao Zheng* $\dagger$, Yunpeng Zhang, Jie Zhou, Jiwen Lu$\ddagger$

* Equal contribution $\dagger$ Project leader $\ddagger$ Corresponding author

News

• [2024/9/5] We have released an object-centric 3D occupancy prediction model GaussianFormer!
• [2023/11/28] We have released a 3D occupancy world model OccWorld!
• [2023/11/23] Check out our recent SelfOcc for self-supervised 3D occupancy prediction!
• [2023/6/14] Code for Semantic Scene Completion on SemanticKITTI has been released.
• [2023/4/12] Check out OccFormer for a high-performing head for occupancy prediction!
• [2023/3/3] We have reorganized the code to improve readability.
• [2023/2/28] TPVFormer is accepted to CVPR 2023!
• [2023/2/26] See SurroundOcc for a more dense 3D occupancy prediction!
• [2023/2/23] Visualization code has been released.
• [2023/2/16] Paper released on arXiv.
• [2023/2/12] Initial code release.
• [2022/11/20] Demo release.

Demo

A full demo video can be downloaded here.

Introduction

Modern methods for vision-centric autonomous driving perception widely adopt the bird's-eye-view (BEV) representation to describe a 3D scene. Despite its better efficiency than voxel representation, it has difficulty describing the fine-grained 3D structure of a scene with a single plane. To address this, we propose a tri-perspective view (TPV) representation which accompanies BEV with two additional perpendicular planes. We model each point in the 3D space by summing its projected features on the three planes. To lift image features to the 3D TPV space, we further propose a transformer-based TPV encoder (TPVFormer) to obtain the TPV features effectively. We employ the attention mechanism to aggregate the image features corresponding to each query in each TPV plane. Experiments show that our model trained with sparse supervision effectively predicts the semantic occupancy for all voxels. We demonstrate for the first time that using only camera inputs can achieve comparable performance with LiDAR-based methods on the LiDAR segmentation task on nuScenes.

Installation

1. Create conda environment with python version 3.8

2. Install pytorch and torchvision with versions specified in requirements.txt

3. Follow instructions in https://mmdetection3d.readthedocs.io/en/latest/getting_started.html#installation to install mmcv-full, mmdet, and mmsegmentation with versions specified in requirements.txt

4. Install timm, numba and pyyaml with versions specified in requirements.txt

Preparing

1. Download pretrain weights from https://github.com/zhiqi-li/storage/releases/download/v1.0/r101_dcn_fcos3d_pretrain.pth and put it in ckpts/

2. Create soft link from data/nuscenes to your_nuscenes_path

3. Download our generated train/val pickle files and put them in data/ nuscenes_infos_train.pkl https://cloud.tsinghua.edu.cn/f/ede3023e01874b26bead/?dl=1 nuscenes_infos_val.pkl https://cloud.tsinghua.edu.cn/f/61d839064a334630ac55/?dl=1

The dataset should be organized as follows:

TPVFormer/data
    nuscenes                 -    downloaded from www.nuscenes.org
        lidarseg
        maps
        samples
        sweeps
        v1.0-trainval
    nuscenes_infos_train.pkl
    nuscenes_infos_val.pkl

Getting Started

Training

1. Train TPVFormer for lidar segmentation task on A100 with 40G GPU memory.

bash launcher.sh config/tpv_lidarseg.py out/tpv_lidarseg 

2. Train TPVFormer for lidar segmentation task on 3090 with 24G GPU memory.

bash launcher.sh config/tpv_lidarseg_dim64.py out/tpv_lidarseg_dim64

3. Train TPVFormer for 3D semantic occupancy prediction task on 3090 with 24G GPU memory.

bash launcher.sh config/tpv04_occupancy.py out/tpv_occupancy

Evaluation for LiDAR Segmentation Metric

1. We release the weights for 3D semantic occupancy prediction at https://cloud.tsinghua.edu.cn/f/3fbd12101ead4397a0f7/?dl=1, and the weights for LiDAR segmentation at https://cloud.tsinghua.edu.cn/f/f686e97feb18406592d9/?dl=1.

2. Run eval.py to calculate mIoU for lidarseg.

python eval.py --py-config xxxx --ckpt-path xxxx

Visualizations

Read visualization/readme.md for detailed instructions.

Semantic Scene Completion on SemanticKITTI

Please refer to kitti_ssc for more details.

Comparisons with Tesla's Occupancy Network

	Tesla's Occupancy Network	Our TPVFormer
Volumetric Occupancy	Yes	Yes
Occupancy Semantics	Yes	Yes
#Semantics	>= 5	16
Input	8 camera images	6 camera images
Training Supervision	Dense 3D reconstruction	Sparse LiDAR semantic labels
Training Data	~1,440,000,000 frames	28,130 frames
Arbitrary Resolution	Yes	Yes
Video Context	Yes	Not yet
Training Time	~100,000 gpu hours	~300 gpu hours
Inference Time	~10 ms on the Tesla FSD computer	~290 ms on a single A100

Related Projects

Our code is based on BEVFormer and Cylinder3D. Many thanks to them!

Welcome to see SurroundOcc for a more dense 3D occupancy prediction. You can also use SurroundOcc's generated dense occupancy ground truths for training instead of sparse lidar supervision.

Also, remember to check out OccFormer for an effective and efficient transformer encoder-decoder specially designed for occupancy prediction.

Citation

If you find this project helpful, please consider citing the following paper:

@article{huang2023tri,
    title={Tri-Perspective View for Vision-Based 3D Semantic Occupancy Prediction},
    author={Huang, Yuanhui and Zheng, Wenzhao and Zhang, Yunpeng and Zhou, Jie and Lu, Jiwen },
    journal={arXiv preprint arXiv:2302.07817},
    year={2023}
}