(TGRS 2024) OrientedFormer: An End-to-End Transformer-Based Oriented Object Detector in Remote Sensing Images
The Chinese Version is below (中文版在下面).
The paper is officially accepted by IEEE Transactions on Geoscience and Remote Sensing (TGRS 2024).
TGRS paper link https://ieeexplore.ieee.org/document/10669376
arxiv link https://arxiv.org/abs/2409.19648
NEW
-
ICDAR2015 Dataset in MMRotate-1.x
-
ICDAR2015 Metric in MMRotate-1.x
-
ChannelMapperWithGN in MMRotate-1.x
-
RBBoxL1Cost in MMRotate-1.x
-
RotatedIoUCost in MMRotate-1.x
-
TopkHungarianAssigner in MMRotate-1.x
If you like it, please click on star.
Please refer to Installation for more detailed instruction.
Note: Our codes base on the newest version mmrotate-1.x, not mmrotate-0.x.
Note: All of our codes can be found in path './projects/OrientedFormer/'.
DOTA and DIOR-R : Please refer to Preparation for more detailed data preparation.
ICDAR2015 : (1) Download ICDAR2015 dataset from official link. (2) The data structure is as follows:
root
├── icdar2015
│ ├── ic15_textdet_train_img
│ ├── ic15_textdet_train_gt
│ ├── ic15_textdet_test_img
│ ├── ic15_textdet_test_gt1). DIOR-R
bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dior.py 2bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_swin-tiny_q300_layer2_head64_point32_1x_dior.py 2bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_lsk_t_q300_layer2_head64_point32_1x_dior.py 22). DOTA-v1.0
bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.0.py 2bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.0-ms.py 2bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_r101_q300_layer2_head64_point32_1x_dotav1.0.py 2bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_swin-tiny_q300_layer2_head64_point32_1x_dotav1.0.py 23). DOTA-v1.5
bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.5.py 24). DOTA-v2.0
bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav2.0.py 25). ICDAR2015
bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_2x_icdar2015.py 21). DIOR-R
bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dior.py work_dirs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dior/epoch_12.pth 2bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_swin-tiny_q300_layer2_head64_point32_1x_dior.py work_dirs/orientedformer_le90_swin-tiny_q300_layer2_head64_point32_1x_dior/epoch_12.pth 2bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_lsk_t_q300_layer2_head64_point32_1x_dior.py work_dirs/orientedformer_le90_lsk_t_q300_layer2_head64_point32_1x_dior/epoch_12.pth 22). DOTA-v1.0
bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.0.py work_dirs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.0/epoch_12.pth 2bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.0-ms.py work_dirs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.0-ms/epoch_12.pth 2bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_r101_q300_layer2_head64_point32_1x_dotav1.0.py work_dirs/orientedformer_le90_r101_q300_layer2_head64_point32_1x_dotav1.0/epoch_12.pth 2bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_swin-tiny_q300_layer2_head64_point32_1x_dotav1.0.py work_dirs/orientedformer_le90_swin-tiny_q300_layer2_head64_point32_1x_dotav1.0/epoch_12.pth 2Upload results to DOTA official website.
3). DOTA-v1.5
bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.5.py work_dirs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.5/epoch_12.pth 2Upload results to DOTA official website.
4). DOTA-v2.0
bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav2.0.py work_dirs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav2.0/epoch_12.pth 2Upload results to DOTA official website.
5). ICDAR2015
Get result submit.zip
python tools/test.py projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_2x_icdar2015.py work_dirs/orientedformer_le90_r50_q300_layer2_head64_point32_2x_icdar2015/epoch_21.pth
Calculate precision, recall and F-measure. The script.py adapted from official website.
pip install Polygon3
python projects/icdar2015_evaluation/script.py –g=gt.zip –s=submit.zip
1). DOTA-v1.0
| Dataset | Configs | Download | AP_50 | AP75 | mAP | Backbone | lr schd | bs |
|---|---|---|---|---|---|---|---|---|
| DOTA-v1.0 | orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.0.py | Hugging Face | 75.3729 | 46.390216 | 45.0071 | R50 | 12epoch | 2img*2 rtx2080ti |
| DOTA-v1.0 | orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.0-ms.py | Hugging Face | 79.064371 | 57.463 | 51.891899 | R50 | 12epoch | 2img*2 rtx2080ti |
| DOTA-v1.0 | orientedformer_le90_r101_q300_layer2_head64_point32_1x_dotav1.0.py | Hugging Face | 75.915958978 | 49.76108 | 47.11829758 | R101 | 12epoch | 2img*2 rtx2080ti |
| DOTA-v1.0 | orientedformer_le90_swin-tiny_q300_layer2_head64_point32_1x_dotav1.0.py | Hugging Face | 75.8819 | 48.965 | 45.8218 | Swin-T | 12epoch | 2img*2 rtx2080ti |
2). DOTA-v1.5
| Dataset | Configs | Download | AP_50 | Backbone | lr schd | bs |
|---|---|---|---|---|---|---|
| DOTA-v1.5 | orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.5.py | Hugging Face | 67.06 | R50 | 12epoch | 2img*2 rtx2080ti |
Due to the limitation of the length of the paper, all categories of AP for DOTA-1.5 are not available in the paper. Here is a list:
| PL | BD | BR | GTF | SV | LV | SH | TC | BC | ST | SBF | RA | HA | SP | HC | CC | AP50 | AP75 | mAP |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 72.0444 | 77.4554 | 51.2471 | 64.9538 | 64.0453 | 77.0387 | 85.3310 | 90.83699 | 77.31017 | 78.106886 | 56.103059 | 68.776 | 68.140988 | 72.081567 | 58.6135 | 10.855397 | 67.05879 | 39.2845 | 38.78675 |
3). DOTA-v2.0
| Dataset | Configs | Download | AP_50 | Backbone | lr schd | bs |
|---|---|---|---|---|---|---|
| DOTA-v2.0 | orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav2.0.py | Hugging Face | 54.27 | R50 | 12epoch | 2img*2 rtx2080ti |
Due to the limitation of the length of the paper, all categories of AP for DOTA-2.0 are not available in the paper. Here is a list:
| PL | BD | BR | GTF | SV | LV | SH | TC | BC | ST | SBF | RA | HA | SP | HC | CC | airport | helipad | AP50 | AP75 | mAP |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 76.7619 | 51.55655 | 42.3872759 | 60.464159 | 56.482355 | 55.43076 | 66.681058 | 78.6341 | 60.0626 | 69.6894 | 35.0316 | 56.015956 | 51.9962 | 56.20235 | 54.95597 | 24.335 | 67.31572 | 12.9641 | 54.266644 | 28.8561385 | 30.0281367 |
4). DIOR-R
| Dataset | Configs | Download | AP_50 | Backbone | lr schd | bs |
|---|---|---|---|---|---|---|
| DIOR-R | orientedformer_le90_r50_q300_layer2_head64_point32_1x_dior.py | Hugging Face | 67.28 | R50 | 12epoch | 2img*2 rtx2080ti |
| DIOR-R | orientedformer_le90_swin-tiny_q300_layer2_head64_point32_1x_dior.py | Hugging Face | 68.84 | Swin-T | 12epoch | 2img*2 rtx2080ti |
| DIOR-R | orientedformer_le90_lsk_t_q300_layer2_head64_point32_1x_dior.py | Hugging Face | 65.07 | LSK-Net | 12epoch | 2img*2 rtx2080ti |
5). ICDAR-2015
| Dataset | Configs | Download | P | R | F-measure | Backbone | lr schd | bs |
|---|---|---|---|---|---|---|---|---|
| ICDAR2015 | orientedformer_le90_r50_q300_layer2_head64_point32_2x_icdar2015.py | Hugging Face | 85.3 | 74.2 | 79.4 | R50 | 24epoch | 2img*2 rtx2080ti |
“Instead of citing an article with closed or falsely advertised open-source code, it is preferable to cite an article with genuine open-source code.”
@ARTICLE{10669376,
author={Zhao, Jiaqi and Ding, Zeyu and Zhou, Yong and Zhu, Hancheng and Du, Wen-Liang and Yao, Rui and El Saddik, Abdulmotaleb},
journal={IEEE Transactions on Geoscience and Remote Sensing},
title={OrientedFormer: An End-to-End Transformer-Based Oriented Object Detector in Remote Sensing Images},
year={2024},
volume={62},
number={},
pages={1-16},
keywords={Encoding;Object detection;Proposals;Detectors;Remote sensing;Current transformers;Position measurement;End-to-end detectors;oriented object detection;positional encoding (PE);remote sensing;transformer},
doi={10.1109/TGRS.2024.3456240}}
(TGRS 2024) OrientedFormer: An End-to-End Transformer-Based Oriented Object Detector in Remote Sensing Images
论文被IEEE Transactions on Geoscience and Remote Sensing (TGRS 2024) 接受。
TGRS官方论文链接 https://ieeexplore.ieee.org/document/10669376
arxiv link https://arxiv.org/abs/2409.19648
新特性：
-
数据集工具： ICDAR2015 Dataset in MMRotate-1.x
-
数据集工具： ICDAR2015 Metric in MMRotate-1.x
-
不用FPN，用channelmapper： ChannelMapperWithGN in MMRotate-1.x
-
端到端分配的L1代价矩阵： RBBoxL1Cost in MMRotate-1.x
-
端到端分配的IoU代价矩阵： RotatedIoUCost in MMRotate-1.x
-
Topk匈牙利匹配： TopkHungarianAssigner in MMRotate-1.x
如果喜欢，请点一点小星星收藏。
“与其引用不开源代码、假开源代码的文章，不如引用真开源代码的文章。”
参考mmrotate-1.x的官方安装教程获取更多安装细节。
注意：代码是基于最新版本的mmrotate-1.x，而不是旧版的mmrotate-0.x。
注意：orientedformer的代码全部在路径'./projects/OrientedFormer/'
DOTA and DIOR-R : 请参考mmrotate-1.x官方数据处理对DOTA和DIOR-R的准备方法。
ICDAR2015 : (1) 从官方网站下载 ICDAR2015 数据集。(2) 数据路径结构如下所示：
root
├── icdar2015
│ ├── ic15_textdet_train_img
│ ├── ic15_textdet_train_gt
│ ├── ic15_textdet_test_img
│ ├── ic15_textdet_test_gt1). DIOR-R
bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dior.py 2bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_swin-tiny_q300_layer2_head64_point32_1x_dior.py 2bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_lsk_t_q300_layer2_head64_point32_1x_dior.py 22). DOTA-v1.0
bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.0.py 2bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.0-ms.py 2bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_r101_q300_layer2_head64_point32_1x_dotav1.0.py 2bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_swin-tiny_q300_layer2_head64_point32_1x_dotav1.0.py 23). DOTA-v1.5
bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.5.py 24). DOTA-v2.0
bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav2.0.py 25). ICDAR2015
bash tools/dist_train.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_2x_icdar2015.py 21). DIOR-R
bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dior.py work_dirs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dior/epoch_12.pth 2bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_swin-tiny_q300_layer2_head64_point32_1x_dior.py work_dirs/orientedformer_le90_swin-tiny_q300_layer2_head64_point32_1x_dior/epoch_12.pth 2bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_lsk_t_q300_layer2_head64_point32_1x_dior.py work_dirs/orientedformer_le90_lsk_t_q300_layer2_head64_point32_1x_dior/epoch_12.pth 22). DOTA-v1.0
bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.0.py work_dirs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.0/epoch_12.pth 2bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.0-ms.py work_dirs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.0-ms/epoch_12.pth 2bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_r101_q300_layer2_head64_point32_1x_dotav1.0.py work_dirs/orientedformer_le90_r101_q300_layer2_head64_point32_1x_dotav1.0/epoch_12.pth 2bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_swin-tiny_q300_layer2_head64_point32_1x_dotav1.0.py work_dirs/orientedformer_le90_swin-tiny_q300_layer2_head64_point32_1x_dotav1.0/epoch_12.pth 2将结果上传 DOTA官方网站。
3). DOTA-v1.5
bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.5.py work_dirs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.5/epoch_12.pth 2将结果上传 DOTA官方网站。
4). DOTA-v2.0
bash tools/dist_test.sh projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav2.0.py work_dirs/orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav2.0/epoch_12.pth 2将结果上传 DOTA官方网站。
5). ICDAR2015
得到结果submit.zip
python tools/test.py projects/OrientedFormer/configs/orientedformer_le90_r50_q300_layer2_head64_point32_2x_icdar2015.py work_dirs/orientedformer_le90_r50_q300_layer2_head64_point32_2x_icdar2015/epoch_21.pth
计算precision, recall 和 F-measure
pip install Polygon3
python projects/icdar2015_evaluation/script.py –g=gt.zip –s=submit.zip
1). DOTA-v1.0
| Dataset | Configs | Download | AP_50 | AP75 | mAP | Backbone | lr schd | bs |
|---|---|---|---|---|---|---|---|---|
| DOTA-v1.0 | orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.0.py | Hugging Face | 75.3729 | 46.390216 | 45.0071 | R50 | 12epoch | 2img*2 rtx2080ti |
| DOTA-v1.0 | orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.0-ms.py | Hugging Face | 79.064371 | 57.463 | 51.891899 | R50 | 12epoch | 2img*2 rtx2080ti |
| DOTA-v1.0 | orientedformer_le90_r101_q300_layer2_head64_point32_1x_dotav1.0.py | Hugging Face | 75.915958978 | 49.76108 | 47.11829758 | R101 | 12epoch | 2img*2 rtx2080ti |
| DOTA-v1.0 | orientedformer_le90_swin-tiny_q300_layer2_head64_point32_1x_dotav1.0.py | Hugging Face | 75.8819 | 48.965 | 45.8218 | Swin-T | 12epoch | 2img*2 rtx2080ti |
2). DOTA-v1.5
| Dataset | Configs | Download | AP_50 | Backbone | lr schd | bs |
|---|---|---|---|---|---|---|
| DOTA-v1.5 | orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav1.5.py | Hugging Face | 67.06 | R50 | 12epoch | 2img*2 rtx2080ti |
由于论文长度的限制，论文中没有DOTA-1.5的所有类别的AP，这里列出：
| PL | BD | BR | GTF | SV | LV | SH | TC | BC | ST | SBF | RA | HA | SP | HC | CC | AP50 | AP75 | mAP |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 72.0444 | 77.4554 | 51.2471 | 64.9538 | 64.0453 | 77.0387 | 85.3310 | 90.83699 | 77.31017 | 78.106886 | 56.103059 | 68.776 | 68.140988 | 72.081567 | 58.6135 | 10.855397 | 67.05879 | 39.2845 | 38.78675 |
3). DOTA-v2.0
| Dataset | Configs | Download | AP_50 | Backbone | lr schd | bs |
|---|---|---|---|---|---|---|
| DOTA-v2.0 | orientedformer_le90_r50_q300_layer2_head64_point32_1x_dotav2.0.py | Hugging Face | 54.27 | R50 | 12epoch | 2img*2 rtx2080ti |
由于论文长度的限制，论文中没有DOTA-2.0的所有类别的AP，这里列出：
| PL | BD | BR | GTF | SV | LV | SH | TC | BC | ST | SBF | RA | HA | SP | HC | CC | airport | helipad | AP50 | AP75 | mAP |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 76.7619 | 51.55655 | 42.3872759 | 60.464159 | 56.482355 | 55.43076 | 66.681058 | 78.6341 | 60.0626 | 69.6894 | 35.0316 | 56.015956 | 51.9962 | 56.20235 | 54.95597 | 24.335 | 67.31572 | 12.9641 | 54.266644 | 28.8561385 | 30.0281367 |
4). DIOR-R
| Dataset | Configs | Download | AP_50 | Backbone | lr schd | bs |
|---|---|---|---|---|---|---|
| DIOR-R | orientedformer_le90_r50_q300_layer2_head64_point32_1x_dior.py | Hugging Face | 67.28 | R50 | 12epoch | 2img*2 rtx2080ti |
| DIOR-R | orientedformer_le90_swin-tiny_q300_layer2_head64_point32_1x_dior.py | Hugging Face | 68.84 | Swin-T | 12epoch | 2img*2 rtx2080ti |
| DIOR-R | orientedformer_le90_lsk_t_q300_layer2_head64_point32_1x_dior.py | Hugging Face | 65.07 | LSK-Net | 12epoch | 2img*2 rtx2080ti |
5). ICDAR-2015
| Dataset | Configs | Download | P | R | F-measure | Backbone | lr schd | bs |
|---|---|---|---|---|---|---|---|---|
| ICDAR2015 | orientedformer_le90_r50_q300_layer2_head64_point32_2x_icdar2015.py | Hugging Face | 85.3 | 74.2 | 79.4 | R50 | 24epoch | 2img*2 rtx2080ti |
“与其引用不开源代码、假开源代码的文章，不如引用真开源代码的文章。”
@ARTICLE{10669376,
author={Zhao, Jiaqi and Ding, Zeyu and Zhou, Yong and Zhu, Hancheng and Du, Wen-Liang and Yao, Rui and El Saddik, Abdulmotaleb},
journal={IEEE Transactions on Geoscience and Remote Sensing},
title={OrientedFormer: An End-to-End Transformer-Based Oriented Object Detector in Remote Sensing Images},
year={2024},
volume={62},
number={},
pages={1-16},
keywords={Encoding;Object detection;Proposals;Detectors;Remote sensing;Current transformers;Position measurement;End-to-end detectors;oriented object detection;positional encoding (PE);remote sensing;transformer},
doi={10.1109/TGRS.2024.3456240}}