- Release the paper on arXiv.
- Release the complete code.
- Release the checkpoints.
- Release training configurations and model checkpoints on 5 additional datasets.
!!! We have open-sourced the code and model checkpoints. Note that the performance reproduced with the refactored code is fully aligned with, and in some cases shows a slight improvement over, the results reported in the paper.
| methods | e2e | mAp | config_file | log | model | |
|---|---|---|---|---|---|---|
| PLA | Y | 56.43 | point2rbox_v3-1x-dotav1-0 | 20251022_160639 | dota_evaluation_results_feedback_of_task1 | epoch_12.pth |
| Point2RBox-v3 | Y | 61.38 | point2rbox_v3-1x-dotav1-0 | 20251022_160639 | dota_evaluation_results_feedback_of_task1 | epoch_12.pth |
| Point2RBox-v3 | N | 67.24 | rotated-fcos-1x-dotav1-0-using-pseudo | 20251028_191527 | dota_evaluation_results_feedback_of_task1 | epoch_12.pth |
| method | dataset | e2e | mAp | config_file | log | text info | model |
|---|---|---|---|---|---|---|---|
| Point2RBox-v3 | STAR | Y | 16.20 | point2rbox_v3-1x-star | 20251031_181448 | results-tab-409093 | epoch_12.pth |
| Point2RBox-v3 | STAR | N | 19.60 | rotated-fcos-1x-star-using-pseudo | 20251102_001001 | results-tab-350178 | epoch_12.pth |
| Point2RBox-v3 | DIOR | Y | 41.70 | point2rbox_v3-1x-dior | 20251031_202113 | x | epoch_12.pth |
| Point2RBox-v3 | DIOR | N | 46.60 | rotated-fcos-1x-dior-using-pseudo | 20251101_135239 | x | epoch_12.pth |
| Point2RBox-v3 | DOTAV1-5 | Y | 49.08 | point2rbox_v3-1x-dotav1-5 | 20251108_172545 | DOTA-v1.5_Evaluation_Results_Feedback_of_Task1 | epoch_12.pth |
- Environment Setup — Step-by-step installation guide with pinned package versions. Please read this carefully before proceeding.
- Data Preparation — Download links, directory structures, and splitting instructions for all supported datasets.
Interactive launcher scripts are provided for convenience. They automatically detect available GPUs, let you select a config, and build the full command:
# Training (interactive)
bash train.sh
# Testing (interactive)
bash test.shBy default, the scripts activate the point2rbox-v3 conda environment. To use a different environment name:
CONDA_ENV=your_env_name bash train.shYou can also run training/testing directly without the launcher:
# Train on a single GPU
CUDA_VISIBLE_DEVICES=0 python tools/train.py configs/point2rbox_v3/point2rbox_v3-1x-dotav1-0.py
# Test with a checkpoint
CUDA_VISIBLE_DEVICES=0 python tools/test.py configs/point2rbox_v3/point2rbox_v3-1x-dotav1-0.py work_dirs/point2rbox_v3-1x-dotav1-0/epoch_12.pth
# two-stage training (total 2 steps)## step1: generator pseudo label
CUDA_VISIBLE_DEVICES=0 python tools/test.py configs/point2rbox_v3/point2rbox_v3-pseudo-generator-dotav1-0.py work_dirs/point2rbox_v3-1x-dotav1-0/epoch_12.pth
## step2: start to train
CUDA_VISIBLE_DEVICES=0 python tools/train.py configs/point2rbox_v3/rotated-fcos-1x-dotav1-0-using-pseudo.py
# two-stage test
CUDA_VISIBLE_DEVICES=0 python tools/test.py configs/point2rbox_v3/rotated-fcos-1x-dotav1-0-using-pseudo.py work_dirs/rotated-fcos-1x-dotav1-0-using-pseudo/epoch_12.pthDOTA v1.0/1.5/2.0 evaluation website: A Large-Scale Benchmark and Challenges for Object Detection in Aerial Images
STAR evaluation website: OBJECT DETECTION IN STAR DATASET
The DIOR and RSAR datasets: evaluated locally without official websites. It should be noted that on the official GitHub repository of RSAR, users have reported missing annotation files for the eval and test datasets. If you encounter a similar issue, you can contact the official authors @zhasion. On the other hand, some other users @wokaikaixinxin have uploaded the files to: modelscope-RSAR
If you want to quickly enter the remote sensing field or follow our method, we provide download links for the preprocessed dataset files here.
- Visual Comparison & Radar Evaluation.
- An Overview of Point2RBox-v3 and Pipeline.
- The process of Progressive Label Assignment (PLA).
- Comparison between watershed and SAM masks on DOTA-v1.0.
- Detection performance of all categories and the mean AP50 on the DOTA-v1.0
- AP$_{50}$ comparisons on the DOTA-v1.0/1.5/2.0, DIOR, STAR, and RSAR datasets.
- AP$_{50}$ comparison on DOTA-v1.0/v1.5 under the partial weakly-supervised setting.
If you have any questions about this paper or code, feel free to email me at zhangteng@sjtu.edu.cn. This ensures I can promptly notice and respond!