ToBL

Code for paper ToBL: Towards Better Labeling for Open World SAR Object Detection.

Requirements

• Linux or macOS with Python ≥ 3.8.
• Install PyTorch ≥ 1.9.0, torchvision, Detectron2, timm, and einops. Recommended detectron2 installation method:

  conda install conda-forge::detectron2
  pip install protobuf==3.20.*
  pip install tensorboard
  pip install black==21.4b2

• Prepare datasets:
  • Download SAR-AIRcraft 1.0 and SARDet-100k.
  • Convert SARDet-100k annotation format using coco_to_voc.py.
  • Move all images to datasets/JPEGImages and annotations to datasets/Annotations.
  • run pip install -r requirements.txt
  • To build object keypoints, run python datasets/build_keypoints.py, then modify config file t1.yaml(SCATTER_FILENAME)

Getting Started

• Training for open world SAR object detection:

  bash run_aircraft.sh
  bash run_sardet100k.sh

  Evaluation for open world object detection:

  bash test_owod.sh

• Visualize the results:

  python demo.py -i LIST_OF_IMAGES  
  # or use fiftyone
  sh core/visualization/vis_fiftyone.sh

• Note that we are using an ImageNet pre-trained backbone. To switch to a DINO pre-trained backbone, please download the model weights and then follow these instructions.

Results

The following results were obtained with four NVIDIA 3090 GPUs, using the checkpoints at this link.

• Open world object detection on SAR-AIRcraft and SARDet-100k:

  

Citation

If you find this code useful, please consider citing:

Acknowledgement

Our implementation is based on OrthogonalDet which uses Detectron2 and Sparse R-CNN. Their contributions have significantly advanced and promoted research in this field.