AdaMixer: A Fast-Converging Query-Based Object Detector
accept to CVPR 2022 as an oral presentation
Ziteng Gao, Limin Wang, Bing Han, Sheng Guo
Nanjing University, MYbank Ant Group
[2022.7.3] Reproduced model checkpoints and logs are available.
[2022.4.4] The code is available now.
[2022.3.31] Code will be released in a few days (not too long). Pre-trained models will take some time to grant the permission of Ant Group to be available online. Please stay tuned or watch this repo for quick information.
To our best knowledge, we are the first to introduce the MLP-Mixer for Object detection. The MLP-Mixer is used in the DETR-like decoder in an adaptive and query-wise manner to enrich the adaptibility to varying objects across images.
AdaMixer enjoys fast convergence speed and reach up to 45.0 AP on COCO val within 12 epochs with only the architectural design improvement. Our method is compatible with other training improvements, like multiple predictions from a query and denosing training, which are expected to improve AdaMixer further (we have not tried yet).
Our AdaMixer does not hunger for extra attention encoders or explicit feature pyramid networks. Instead, we improve the query decoder in DETR-like detectors to keep the architecture as simple, efficient, and strong as possible.
Our code structure follows the MMDetection framework. To get started, please refer to mmdetection doc get_started.md for installation.
Our AdaMixer config file lies in configs/adamixer folder. You can start training our detectors with make targets in Makefile.
The code of a AdaMixer decoder stage is in mmdet/models/roi_heads/bbox_heads/adamixer_decoder_stage.py. The code of the 3D feature space sampling is in mmdet/models/roi_heads/bbox_heads/sampling_3d_operator.py. The code of the adaptive mixing process is in mmdet/models/roi_heads/bbox_heads/adaptive_mixing_operator.py.
NOTE:
mmcv_full==1.3.3
and pytorch>=1.5.0
for correct reproduction (#4, #12).init_weight
methods in AdaptiveSamplingMixing
and AdaptiveMixing
are called for correct initializations AND the initialized weights are not overrided by other methods (some MMCV versions may incur repeated initializations).Checkpoints and logs are available at google drive.
config | detector | backbone | APval | APtest | APval (reprod.) | ckpt (reprod.) | log (reprod.) |
---|---|---|---|---|---|---|---|
config | AdaMixer (1x schedule, 100 queries) | R50 | 42.7 | 42.6 | ckpt | log | |
config | AdaMixer (3x schedule, 300 queries) | R50 | 47.0 | 47.2 | 46.8 | ckpt | log |
config | AdaMixer (3x schedule, 300 queries) | R101 | 48.0 | 48.1 | 48.1 | ckpt | log |
config | AdaMixer (3x schedule, 300 queries) | X101-DCN | 49.5 | 49.3 | 49.7 | ckpt | log |
config | AdaMixer (3x schedule, 300 queries) | Swin-S | 51.3 | 51.3 | on the way |
Special thanks to Zhan Tong for these reproduced models.
If you find AdaMixer useful in your research, please cite us using the following entry:
@inproceedings{adamixer22cvpr,
author = {Ziteng Gao and
Limin Wang and
Bing Han and
Sheng Guo},
title = {AdaMixer: A Fast-Converging Query-Based Object Detector},
booktitle = {{CVPR}},
year = {2022}
}
Thanks to Zhan Tong and Zihua Xiong for their help.
The following begins the original mmdetection README.md file
News: We released the technical report on ArXiv.
Documentation: https://mmdetection.readthedocs.io/
English | 简体中文
MMDetection is an open source object detection toolbox based on PyTorch. It is a part of the OpenMMLab project.
The master branch works with PyTorch 1.3+. The old v1.x branch works with PyTorch 1.1 to 1.4, but v2.0 is strongly recommended for faster speed, higher performance, better design and more friendly usage.
Modular Design
We decompose the detection framework into different components and one can easily construct a customized object detection framework by combining different modules.
Support of multiple frameworks out of box
The toolbox directly supports popular and contemporary detection frameworks, e.g. Faster RCNN, Mask RCNN, RetinaNet, etc.
High efficiency
All basic bbox and mask operations run on GPUs. The training speed is faster than or comparable to other codebases, including Detectron2, maskrcnn-benchmark and SimpleDet.
State of the art
The toolbox stems from the codebase developed by the MMDet team, who won COCO Detection Challenge in 2018, and we keep pushing it forward.
Apart from MMDetection, we also released a library mmcv for computer vision research, which is heavily depended on by this toolbox.
The mmdetection project is released under the Apache 2.0 license.
v2.12.0 was released in 01/05/2021. Please refer to changelog.md for details and release history. A comparison between v1.x and v2.0 codebases can be found in compatibility.md.
Results and models are available in the model zoo.
Supported backbones:
Supported methods:
Some other methods are also supported in projects using MMDetection.
Please refer to get_started.md for installation.
Please see get_started.md for the basic usage of MMDetection. We provide colab tutorial, and full guidance for quick run with existing dataset and with new dataset for beginners. There are also tutorials for finetuning models, adding new dataset, designing data pipeline, customizing models, customizing runtime settings and useful tools.
Please refer to FAQ for frequently asked questions.
We appreciate all contributions to improve MMDetection. Please refer to CONTRIBUTING.md for the contributing guideline.
MMDetection is an open source project that is contributed by researchers and engineers from various colleges and companies. We appreciate all the contributors who implement their methods or add new features, as well as users who give valuable feedbacks. We wish that the toolbox and benchmark could serve the growing research community by providing a flexible toolkit to reimplement existing methods and develop their own new detectors.
If you use this toolbox or benchmark in your research, please cite this project.
@article{mmdetection,
title = {{MMDetection}: Open MMLab Detection Toolbox and Benchmark},
author = {Chen, Kai and Wang, Jiaqi and Pang, Jiangmiao and Cao, Yuhang and
Xiong, Yu and Li, Xiaoxiao and Sun, Shuyang and Feng, Wansen and
Liu, Ziwei and Xu, Jiarui and Zhang, Zheng and Cheng, Dazhi and
Zhu, Chenchen and Cheng, Tianheng and Zhao, Qijie and Li, Buyu and
Lu, Xin and Zhu, Rui and Wu, Yue and Dai, Jifeng and Wang, Jingdong
and Shi, Jianping and Ouyang, Wanli and Loy, Chen Change and Lin, Dahua},
journal= {arXiv preprint arXiv:1906.07155},
year={2019}
}