RTS3D: Real-time Stereo 3D Detection from 4D Feature-Consistency Embedding Space for Autonomous Driving (AAAI2021).
RTS3D is efficiency and accuracy stereo 3D object detection method for autonomous driving.
Introduction
RTS3D is the first true real-time system (FPS>24) for stereo image 3D detection meanwhile achieves 10% improvement in average precision comparing with the previous state-of-the-art method. RTS3D only require RGB images without synthetic data, instance segmentation, CAD model, or depth generator.
Highlights
- Fast: 33 FPS of single image test speed in KITTI benchmark with 384*1280 resolution
- Accuracy: SOTA on the KITTI benchmark.
- Anchor Free: No 2D or 3D anchor are reauired
- Easy to deploy: RTS3D uses conventional convolution operations and MLP, so it is very easy to deploy and accelerate.
RTS3D Baseline and Model Zoo
All experiments are tested with Ubuntu 16.04, Pytorch 1.0.0, CUDA 9.0, Python 3.6, single NVIDIA 2080Ti
IoU Setting 1: Car IoU > 0.5, Pedestrian IoU > 0.25, Cyclist IoU > 0.25
IoU Setting 2: Car IoU > 0.7, Pedestrian IoU > 0.5, Cyclist IoU > 0.5
- Training on KITTI train split and evaluation on val split.
- FCE Space Resolution: 10 * 10 * 10
- Model: (Google Drive), (Baidu Cloud 提取码:k4uk)
Class | Iteration | FPS | AP BEV IoU Setting1 | AP 3D IoU Setting1 | AP BEV IoU Setting2 | AP 3D IoU Setting2 |
---|---|---|---|---|---|---|
- | - | - | Easy / Moderate / Hard | Easy / Moderate / Hard | Easy / Moderate / Hard | Easy / Moderate / Hard |
Car- Recall-11 | 1 | 90.9 | 89.83, 77.05, 68.28 | 89.27, 70.12, 61.17 | 73.20, 53.62, 46.44 | 60.87, 42.38, 36.44 |
Car- Recall-40 | 1 | 90.9 | 92.92, 76.17, 66.62 | 90.35, 71.37, 63.52 | 78.12, 54.75, 47.09 | 60.34, 39.32, 32.97 |
Car- Recall-11 | 2 | 45.5 | 90.41, 78.70, 70.03 | 90.26, 77.23, 68.28 | 76.56, 56.46, 48.20 | 63.65, 44.50, 37.48 |
Car- Recall-40 | 2 | 45.5 | 95.75, 79.61, 69.69 | 93.57, 76.64, 66.72 | 78.12, 54.75, 47.09 | 63.99, 41.78, 34.96 |
- Training on KITTI train split and evaluation on val split.
- FCE Space Resolution: 10 * 10 * 10
- Recall split: 11
- Iteration: 2
- Model: (Google Drive), (Baidu Cloud 提取码:4t4u)
Class | AP BEV IoU Setting1 | AP 3D IoU Setting1 | AP BEV IoU Setting2 | AP 3D IoU Setting2 |
---|---|---|---|---|
- | Easy / Moderate / Hard | Easy / Moderate / Hard | Easy / Moderate / Hard | Easy / Moderate / Hard |
Car | 90.18, 78.46, 69.76 | 89.88, 76.64, 67.86 | 74.95, 54.07, 46.78 | 58.50, 39.74, 34.83 |
Pedestrian | 57.12, 48.82, 40.88 | 56.36, 48.29, 40.22 | 32.16, 26.31, 21.28 | 26.95, 20.77, 19.74 |
Cyclist | 54.48, 35.78, 30.80 | 53.86, 30.90, 30.52 | 33.59, 20.80, 20.14 | 31.05, 20.26, 18.93 |
Installation
Please refer to INSTALL.md
Dataset preparation
Please download the official KITTI 3D object detection dataset and organize the downloaded files as follows:
KM3DNet
├── kitti_format
│ ├── data
│ │ ├── kitti
│ │ | ├── annotations
│ │ │ ├── calib /000000.txt .....
│ │ │ ├── image(left[0-7480] right[7481-14961] input augmentatiom)
│ │ │ ├── label /000000.txt .....
| | | ├── train.txt val.txt trainval.txt
│ │ │ ├── mono_results /000000.txt .....
├── src
├── demo_kitti_format
├── readme
├── requirements.txt
Getting Started
Please refer to GETTING_STARTED.md to learn more usage about this project.
Acknowledgement
License
RTS3D is released under the MIT License (refer to the LICENSE file for details). Portions of the code are borrowed from, CenterNet, iou3d and kitti_eval (KITTI dataset evaluation). Please refer to the original License of these projects (See NOTICE).
Citation
If you find this project useful for your research, please use the following BibTeX entry.
@misc{2012.15072,
Author = {Peixuan Li, Shun Su, Huaici Zhao},
Title = {RTS3D: Real-time Stereo 3D Detection from 4D Feature-Consistency Embedding Space for Autonomous Driving},
Year = {2020},
Eprint = {arXiv:2012.15072},
}