This is the official PyTorch implementation of our NeurIPS 2023 paper:
HASSOD: Hierarchical Adaptive Self-Supervised Object Detection
[Project Page] [Paper-arXiv] [Paper-OpenReview] [Video-YouTube] [Video-Bilibili]
Shengcao Cao, Dhiraj Joshi, Liang-Yan Gui, Yu-Xiong Wang
HASSOD is a fully self-supervised approach for object detection and instance segmentation, demonstrating a significant improvement over the previous state-of-the-art methods by discovering a more comprehensive range of objects. Moreover, HASSOD understands the part-to-whole object composition like humans do, while previous methods cannot. Notably, we improve class-agnostic Mask AR from 20.2 to 22.5 on LVIS, and from 17.0 to 26.0 on SA-1B.
To use our code and reproduce the results, please follow these detailed documents step by step:
- Preparation: Prepare the environment, data, and pre-trained models
- Reproduction: Produce pseudo-labels and train the object detector (download links included for our pseudo-labels and model)
- Demo: Once the preparation is finished, you can try out the demo code and test our model on any image.
Our code is developed based on the following repositories:
We greatly appreciate their open-source work!
This project is released under the Apache 2.0 license. Other codes from open source repository follows the original distributive licenses.
If you find our research interesting or use our code, data, or model in your research, please consider citing our work.
@inproceedings{cao2023hassod,
title={{HASSOD}: Hierarchical Adaptive Self-Supervised Object Detection},
author={Cao, Shengcao and Joshi, Dhiraj and Gui, Liangyan and Wang, Yu-Xiong},
booktitle={NeurIPS},
year={2023}
}