ZBS: Zero-shot Background Subtraction via Instance-level Background Modeling and Foreground Selection
This repository is an official implementation of the ZBS. ZBS fully utilizes the advantages of zero-shot object detection to build the open-vocabulary instance-level background model. It can detect most of the categories in the real world and can detect the unseen foreground categories outside the pre-defined categories. ZBS achieves remarkably 4.70% F-Measure improvements over state-of-the-art unsupervised methods.
-
The first unsupervised zero-shot background subtraction.
-
The first background subtraction method based on an instance-level background model.
-
Detects any class given class names (using Detic).
-
State-of-the-art results on CDnet 2014 dataset compared with other unsupervised background subtraction method.
See GET_STARTED.md.
Overall and per-category F-Measure comparison of different Unsupervised BGS methods on the CDnet 2014 dataset.
Unsupervised BGS | baseline | camjitt | dynbg | intmot | shadow | thermal | badwea | lowfr | night | PTZ | turbul | Overall |
---|---|---|---|---|---|---|---|---|---|---|---|---|
PAWCS | 0.9397 | 0.8137 | 0.8938 | 0.7764 | 0.8913 | 0.8324 | 0.8152 | 0.6588 | 0.4152 | 0.4615 | 0.6450 | 0.7403 |
SuBSENSE | 0.9503 | 0.8152 | 0.8177 | 0.6569 | 0.8986 | 0.8171 | 0.8619 | 0.6445 | 0.5599 | 0.3476 | 0.7792 | 0.7408 |
WisenetMD | 0.9487 | 0.8228 | 0.8376 | 0.7264 | 0.8984 | 0.8152 | 0.8616 | 0.6404 | 0.5701 | 0.3367 | 0.8304 | 0.7535 |
SWCD | 0.9214 | 0.7411 | 0.8645 | 0.7092 | 0.8779 | 0.8581 | 0.8233 | 0.7374 | 0.5807 | 0.4545 | 0.7735 | 0.7583 |
SemanticBGS | 0.9604 | 0.8388 | 0.9489 | 0.7878 | 0.9478 | 0.8219 | 0.8260 | 0.7888 | 0.5014 | 0.5673 | 0.6921 | 0.7892 |
RTSS | 0.9597 | 0.8396 | 0.9325 | 0.7864 | 0.9551 | 0.8510 | 0.8662 | 0.6771 | 0.5295 | 0.5489 | 0.7630 | 0.7917 |
RT-SBS-v2 | 0.9535 | 0.8233 | 0.9217 | 0.8946 | 0.9497 | 0.8697 | 0.8279 | 0.7341 | 0.5629 | 0.5808 | 0.7315 | 0.8045 |
ZBS (Ours) | 0.9653 | 0.9545 | 0.9290 | 0.8758 | 0.9765 | 0.8698 | 0.9229 | 0.7433 | 0.6800 | 0.8133 | 0.6358 | 0.8515 |
Overall and per-category result of ZBS on the CDnet 2014 dataset.
Category | Recall | Specificity | PWC | Precision | F-Measure |
---|---|---|---|---|---|
badWea | 0.9049 | 0.9988 | 0.2755 | 0.9439 | 0.9229 |
baseline | 0.9709 | 0.9988 | 0.2237 | 0.9603 | 0.9653 |
camjitt | 0.9543 | 0.9979 | 0.4022 | 0.9554 | 0.9545 |
dynbg | 0.9269 | 0.9996 | 0.0951 | 0.9340 | 0.9290 |
intmot | 0.8254 | 0.9965 | 1.6864 | 0.9481 | 0.8758 |
lowfr | 0.7302 | 0.9988 | 0.3279 | 0.7584 | 0.7433 |
night | 0.6341 | 0.9958 | 1.2477 | 0.7666 | 0.6800 |
PTZ | 0.7490 | 0.9997 | 0.2387 | 0.9223 | 0.8133 |
shadow | 0.9712 | 0.9991 | 0.2097 | 0.9819 | 0.9765 |
thermal | 0.8475 | 0.9954 | 1.1686 | 0.9040 | 0.8698 |
turbul | 0.7286 | 0.9984 | 0.3198 | 0.6075 | 0.6358 |
Overall | 0.8403 | 0.9981 | 0.5632 | 0.8802 | 0.8515 |
If you find this project useful for your research, please consider citing this paper.
@inproceedings{
an2023zbs,
title={{ZBS}: Zero-shot Background Subtraction via instance-level background modeling and foreground selection},
author={Yongqi An and Xu Zhao and Tao Yu and Haiyun Guo and Chaoyang Zhao and Ming Tang and Jinqiao Wang},
booktitle={Conference on Computer Vision and Pattern Recognition 2023},
year={2023},
url={https://openreview.net/forum?id=f-9UZN4GEV}
}
Our repository is mainly built upon Detic.