Awesome
Obj2Seq: Formatting Objects as Sequences with Class Prompt for Visual Tasks
Introduction
This repository is an official implementation of the Obj2Seq. Obj2Seq takes objects as basic units, and regards most object-level visual tasks as sequence generation problems of objects. It first recognizes objects of given categories, and then generates a sequence to describe each of these objects. Obj2Seq is able to flexibly determine input categories and the definition of output sequences to satisfy customized requirements, and be easily extended to different visual tasks.
Obj2Seq: Arxiv | Github | Gitee
Main Results
All results are trained with a ResNet-50 backbone.
Object Detection
| Epochs | Params(M) | $AP$ | Model | |
|---|---|---|---|---|
| DeformableDETR$^\dagger$ | 50 | 40 | 44.6 | model |
| Obj2Seq | 50 | 40 | 45.7 | model |
| + iterative box refine | 50 | 42 | 46.7 | model |
$^\dagger$ We convert official DeformableDETR checkpoint with this script.
Human Pose Estimation
| Epochs | Params(M) | $AP_{box}$ | $AP_{kps}$ | Config/Model | |
|---|---|---|---|---|---|
| Baseline | 50 | 40 | 55.4 | 57.9 | model |
| Obj2Seq | 50 | 40 | 55.4 | 61.2 | model |
| Obj2Seq | 150 | 40 | 58.1 | 65.1 | model |
You may also download these models from BaiduNetdisk.
Instructions
See GET_STARTED.md.
Citation
If you find this project useful for your research, please consider citing this paper.
@inproceedings{
chen2022objseq,
title={Obj2Seq: Formatting Objects as Sequences with Class Prompt for Visual Tasks},
author={Zhiyang Chen and Yousong Zhu and Zhaowen Li and Fan Yang and Wei Li and Haixin Wang and Chaoyang Zhao and Liwei Wu and Rui Zhao and Jinqiao Wang and Ming Tang},
booktitle={Advances in Neural Information Processing Systems},
editor={Alice H. Oh and Alekh Agarwal and Danielle Belgrave and Kyunghyun Cho},
year={2022},
url={https://openreview.net/forum?id=cRNl08YWRKq}
}
Acknowledgement
Our repository is mainly built upon DETR, Deformable-DETR and Anchor-DETR. We also refer
- ASL, Query2Label for multi-label classification.
- CLIP, Detic for class-vector generation.
- maskrcnn-benchmark for the dataset with keypoint annotations.
- Swin-Transformer for configs and the swin backbone.