DA-Ada: Learning Domain-Aware Adapter for Domain Adaptive Object Detection

The official implementation of DA-Ada: Learning Domain-Aware Adapter for Domain Adaptive Object Detection (Openreview).

This codebase is based on RegionCLIP.

1. Put your dataset at './datasets/your_dataset'. Please follow the format of Pascal Voc. For example:

• dataset
  • cityscapes_voc
    • VOC2007
      • Annotations
      • ImageSets
      • JPEGImages
  • foggy_cityscapes_voc
    • VOC2007
      • Annotations
      • ImageSets
      • JPEGImages

2. Put your pre-trained VLM model at somewhere you like, for example, './ckpt', and edit the MODEL.WEIGHTS in train_da_ada_c2f.sh.
3. Following RegionCLIP, generate class embedding and put it at somewhere you like, and edit the MODEL.CLIP.TEXT_EMB_PATH.
4. Training: train_da_ada_c2f.sh Testing: test_da_ada_c2f.sh