Wenqiang Zhang*, Tianheng Cheng*, Xinggang Wang†, Shaoyu Chen, Qian Zhang, Wenyu Liu
(*: equal contribution, †: corresponding author)
14 June, 2022: Code and models of Featurized Query R-CNN have been released!
The query mechanism introduced in the DETR method is changing the paradigm of object detection and recently there are many query-based methods have obtained strong object detection performance. However, the current query-based detection pipelines suffer from the following two issues. Firstly, multi-stage decoders are required to optimize the randomly initialized object queries, incurring a large computation burden. Secondly, the queries are fixed after training, leading to unsatisfying generalization capability. To remedy the above issues, we present featurized object queries predicted by a query generation network in the well-established Faster R-CNN framework and develop a Featurized Query R-CNN. Extensive experiments on the COCO dataset show that our Featurized Query R-CNN obtains the best speed-accuracy trade-off among all R-CNN detectors, including the recent state-of-the-art Sparse R-CNN detector.
Our methods are based on detectron2, please refer to here for more details.
Install the detectron2:
git clone https://github.com/facebookresearch/detectron2.git
python setup.py build develop
For training, run:
python train_net.py --config-file <config-file> --num-gpus <num-gpus>
| Model | Backbone | Epoch | AP | FPS | Weights |
|---|---|---|---|---|---|
| Featurized QR-CNN (100 Queries) | ResNet-50 | 36 | 41.3 | 26 | Google Drive |
| Cascade Featurized QR-CNN (100 Queries) | ResNet-50 | 36 | 43.0 | 24 | Google Drive |
| Cascade Featurized QR-CNN (300 Queries) | ResNet-50 | 36 | 44.6 | 24 | Google Drive |
| Cascade Featurized QR-CNN (100 Queries) | ResNet-101 | 36 | 43.9 | 18 | Google Drive |
| Cascade Featurized QR-CNN (300 Queries) | ResNet-101 | 36 | 45.8 | 17 | Google Drive |
- The speed is tested on a single RTX 2080Ti GPU on COCO val set.
- If you have trouble accessing the models in Google Drive, we also provide the models in BaiduPan for you (Password: n91e).
Our implementation is based on detectron2 and Sparse R-CNN, we thank for their open-source code.
Featurized Query R-CNN is released under the MIT Licence.
If you find Featurized Query R-CNN is useful in your research or applications, please consider giving us a star 🌟 and citing it by the following BibTeX entry.
@article{FeaturizedQR-CNN,
title={Featurized Query R-CNN},
author={Zhang, Wenqiang and Cheng, Tianheng and Wang, Xinggang and Chen, Shaoyu and Zhang, Qian and Liu, Wenyu},
journal={arXiv preprint arXiv:2206.04584},
year={2022}
}