CIA-SSD: Confident IoU-Aware Single Stage Object Detector From Point Cloud (AAAI 2021) [Paper]
Currently state-of-the-art single-stage object detector from point cloud on KITTI Benchmark, running with 32FPS.
Authors: Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, Chi-Wing Fu.
Val Split (11 recall points):
Car AP:98.85, 90.20, 89.58
bev AP:90.51, 88.86, 87.95
3d AP:90.00, 79.86, 78.83
aos AP:98.77, 89.99, 89.24
Car AP(Average Precision)@0.70, 0.50, 0.50:
bbox AP:98.85, 90.20, 89.58
bev AP:98.92, 90.29, 89.81
3d AP:99.00, 90.22, 89.70
aos AP:98.77, 89.99, 89.24
Test Split: Submission link
You may download the pre-trained model here, which is trained on the train split (3712 samples).
The pipeline of our proposed Confident IoU-Aware Single-Stage object Detector (CIA-SSD). First, we encode the input point cloud (a) with a sparse convolutional network denoted by SPConvNet (b), followed by our spatial-semantic feature aggregation (SSFA) module (c) for robust feature extraction, in which an attentional fusion module (d) is adopted to adaptively fuse the spatial and semantic features. Then, the multi-task head (e) realizes the object classification and localization, with our introduced confidence function (CF) for confidence rectification. In the end, we further formulate the distance-variant IoU-weighted NMS (DI-NMS) for post-processing.
$ git clone https://github.com/Vegeta2020/CIA-SSD.git
$ cd ./CIA-SSD/det3d/core/iou3d
$ python setup.py install
$ cd ./CIA-SSD
$ python setup.py build developPlease follow Det3D for installation of other related packages and data preparation.
Configure the model in
$ /CIA-SSD/examples/second/configs/kitti_car_vfev3_spmiddlefhd_rpn1_mghead_syncbn.pyPlease use our code to generate ground truth data:
$ python ./CIA-SSD/tools/create_data.pyTrain the CIA-SSD:
$ cd ./CIA-SSD/tools
$ python train.py # Single GPU
$ python -m torch.distributed.launch --nproc_per_node=4 train.py # Multiple GPUEvaluate the CIA-SSD:
$ cd ./CIA-SSD/tools
$ python test.pyIf you find this work useful in your research, please star our repository and consider citing:
@inproceedings{zheng2020ciassd,
title={CIA-SSD: Confident IoU-Aware Single-Stage Object Detector From Point Cloud},
author={Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, Chi-Wing Fu},
booktitle={AAAI},
year={2021}
}
This codebase is released under the Apache 2.0 license.
Our code are mainly based on Det3D, thanks for their contributions! We also thank for the reviewers's valuable comments on this paper.
If you have any question or suggestion about this repo, please feel free to contact me (zheng-w10@foxmail.com)