A DeNoising FPN with Transformer R-CNN for Tiny Object Detection

A PyTorch implementation and pretrained models for DNTR (DeNoising Transformer R-CNN). We present DN-FPN, a plug-in that suppresses noise generated during the fusion of FPNs. In addition, we renalvate the standard R-CNN to consist of a transformer structure, namely Trans R-CNN.(base)

News

[2024/7/1]: DQ-DETR has been accepted by ECCV 2024. 🔥🔥🔥

[2024/5/3]: DNTR has been accepted by TGRS 2024. 🔥🔥🔥

Other Research Paper on Tiny Object Detection

<!-- A DeNoising FPN With Transformer R-CNN for Tiny Object Detection Hou-I Liu and Yu-Wen Tseng and Kai-Cheng Chang and Pin-Jyun Wang and Hong-Han Shuai, and Wen-Huang Cheng IEEE Transactions on Geoscience and Remote Sensing [paper] [code] -->

| Title | Venue | Links | |------|-------------|-------| | DNTR | TGRS 2024 | Paper | code | 中文解读 | | DQ-DETR| ECCV 2024 | Paper | code | 中文解读 |

Installation and Get Started

<!-- Required environments: * Linux * Python 3.6+ * PyTorch 1.3+ * CUDA 9.2+ * GCC 5+ * [MMCV](https://mmcv.readthedocs.io/en/latest/#installation) * [cocoapi-aitod](https://github.com/jwwangchn/cocoapi-aitod) -->

Our model is based on mmdet-aitod and MMDetection.

<!-- This implementation is based on [MMDetection 2.24.1](https://github.com/open-mmlab/mmdetection). Assume that your environment has satisfied the above requirements, -->

Please follow the following steps for installation.

git clone https://github.com/hoiliu-0801/DNTR.git
cd DNTR/mmdet-dntr
# Installation
sh install.sh

Get Started with single GPU

Training DNTR, for example :

python tools/train.py configs/aitod-dntr/aitod_DNTR_mask.py

Testing DNTR, for example :

python tools/test.py configs/aitod-dntr/aitod_DNTR_mask.py

Performance

Table 1. Training Set: AI-TOD trainval set, Testing Set: AI-TOD test set, 36 epochs, where FRCN, DR denotes Faster R-CNN and DetectoRS, respectively. |Method | Backbone | mAP | AP₅₀ | AP₇₅ |AP_vt | AP_t | AP_s | AP_m | |:---:|:---:|:---:|:---:|:---:|:---:|:---: |:---: |:---: | FRCN | R-50 | 11.1 | 26.3 | 7.6 | 0.0 | 7.2 | 23.3 | 33.6 | ATSS | R-50 | 12.8 | 30.6 | 8.5 | 1.9 | 11.6 | 19.5 | 29.2 | ATSS w/ DN-FPN | R-50 | 17.9 | 41.0 | 12.9 | 3.7 | 16.4 | 25.3 | 35.0 | NWD-RKA | R-50 | 23.4 | 53.5 | 16.8 | 8.7 | 23.8 | 28.5 | 36.0 | DNTR | R-50 | 26.2 | 56.7 | 20.2 | 12.8 | 26.4 | 31.0 | 37.0 | DNTR (New) | R-50 | 27.2 | 56.3 | 21.8 | 15.2 | 27.4 | 31.9 | 38.5 |

Table 2. Training Set: Visdrone train set, Validation Set: Visdrone val set, 12 epochs, |Method | Backbone |AP| AP₅₀ | AP₇₅ | |:---:|:---:|:---:|:---:|:---:| DNTR | R-50 | 34.4 | 57.9 | 35.3 | UFPMP w/o DN-FPN| R-50 | 36.6 | 62.4 | 36.7 | UFPMP w/ DN-FPN | R-50 | 37.8 | 62.7 | 38.6 |

AI-TOD-v1 and AI-TOD-v2 Datasets (Don’t forget to leave us a ⭐)

• Step 1: Download the datasets from the below link.

https://drive.google.com/drive/folders/1CowS5BrujefWQxxlmOFfUuLOfUUm8w6U?usp=sharing

Pretrained Weights of AI-TOD-v1 and AI-TOD-v2.

https://drive.google.com/drive/folders/1i0mYPQ3Cz_k4iAIvSwecwpWMX_wivxzY

Note

If you want to run other baseline method with DN-FPN, please replace /mmdet/models/detectors/two_stage_ori.py with mmdet/models/detectors/two_stage.py.

For example: Faster R-CNN: python tools/train.py configs/aitod-dntr/aitod_faster_r50_dntr_1x.py.

Citation

@ARTICLE{10518058,
  author={Liu, Hou-I and Tseng, Yu-Wen and Chang, Kai-Cheng and Wang, Pin-Jyun and Shuai, Hong-Han and Cheng, Wen-Huang},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={A DeNoising FPN With Transformer R-CNN for Tiny Object Detection}, 
  year={2024},
  volume={62},
  number={},
  pages={1-15},
}

@InProceedings{huang2024dq,
author={Huang, Yi-Xin and Liu, Hou-I and Shuai, Hong-Han and Cheng, Wen-Huang},
title={DQ-DETR: DETR with Dynamic Query for Tiny Object Detection},
booktitle={European Conference on Computer Vision},
pages={290--305},
year={2025},
organization={Springer}
}