This repository introduces a few image corruptions for two novel benchmarks to evaluate models' robustness in aerial object detection. All these corruptions are applied to DOTA-v1.0 test set and not used as data augmentation strategies in models' training phase.
The first benchmark encompasses 19 prevalent corruptions. For more detailed information, you may refer to the paper on the original corruption package authored by Hendrycks and Dietterich: Benchmarking neural network robustness to common corruptions and perturbations.
The second benchmark focuses on cloud-corrupted images—a phenomenon uncommon in natural pictures yet frequent in aerial photography. Process 1 represents "Cloud Self-Subtraction" and process 2 represents "Cloud Addition-to-Scene". The detailed principle of this data processing method can be referred to Cloudy Image Arithmetic: A Cloudy Scene Synthesis Paradigm With an Application to Deep-Learning-Based Thin Cloud Removal.
Note that, unless explicitly stated, the backbone of all models is ResNet-50.
| Method | Reference | mPC | rPC (%) | rPCclouds (%) | ||
|---|---|---|---|---|---|---|
| Rotated Faster R-CNN | Ren et al. | 73.4 | 38.7 | 52.7 | 58.5 | 79.7 |
| RoI Transformer | Ding et al. | 76.1 | 39.7 | 52.1 | 60.0 | 78.9 |
| Oriented R-CNN | Xie et al. | 75.7 | 40.4 | 53.4 | 60.6 | 80.1 |
| ReDet | Han et al. | 76.7 | 45.6 | 59.5 | 66.2 | 86.3 |
| Rotated RetinaNet | Lin et al. | 68.4 | 37.1 | 54.2 | 55.1 | 80.6 |
| Rotated FCOS | Tian et al. | 71.3 | 38.6 | 54.2 | 57.5 | 80.7 |
| R3Det | Yang et al. | 69.8 | 37.6 | 53.8 | 56.7 | 81.2 |
| S2A-Net | Han et al. | 73.9 | 39.5 | 53.4 | 59.3 | 80.2 |
| RoI Transformer (backbone=ConvNeXt-T) | Liu et al. | 75.0 | 47.5 | 63.3 | 64.5 | 86.0 |
| RoI Transformer (backbone=Swin-T) | Liu et al. | 77.5 | 43.1 | 55.6 | 62.8 | 81.1 |
| RoI Transformer (backbone=Swin-S) | Liu et al. | 77.1 | 44.3 | 57.4 | 63.3 | 82.1 |
| RoI Transformer (backbone=Swin-B) | Liu et al. | 77.7 | 44.8 | 57.6 | 64.8 | 83.5 |
| RoI Transformer (backbone=Swin-L) | Liu et al. | 77.6 | 47.5 | 61.2 | 66.7 | 85.9 |
| RoI Transformer (augmentation=RandomRotate) | 76.4 | 40.9 | 53.6 | 61.3 | 80.2 | |
| RoI Transformer (augmentation=Mosaic) | Bochkovskiy et al. | 74.4 | 38.8 | 52.2 | 59.6 | 80.1 |
If you make use of the data in DOTA-C, please cite our following paper:
@inproceedings{xia2018dota,
title={DOTA: A large-scale dataset for object detection in aerial images},
author={Xia, Gui-Song and Bai, Xiang and Ding, Jian and Zhu, Zhen and Belongie, Serge and Luo, Jiebo and Datcu, Mihai and Pelillo, Marcello and Zhang, Liangpei},
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
pages={3974--3983},
year={2018}
}
@misc{he2023robustness,
title={On the Robustness of Object Detection Models in Aerial Images},
author={Haodong He and Jian Ding and Gui-Song Xia},
year={2023},
eprint={2308.15378},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent