The dataset

[dusongjie]

Could you please provide more information about the data set download, which data set is required for this experiment on the official ACDC website? And if I want to experiment on other data sets, how should the data set be organized? The training images of the target domain have no labels, and the validation images of the target domain have labels. Do I understand this correctly?

[davidjaw]

• Firstly, the dataset structure will be like the followings, while images contain {filename}.jpg or other image format, and labels contain {filename}.txt, and the label txt format is cls, cx, cy, w, h and cx, cy, w, h are normalized value:

  ── dataset root
  ├─train
  │  ├─fog
  │  │  ├─images
  │  │  └─labels
  │  ├─fog_rain
  │  │  ├─images
  │  │  └─labels
  │  ├─night
  │  │  ├─images
  │  │  └─labels
  │  ├─rain
  │  │  ├─images
  │  │  └─labels
  │  └─snow
  │      ├─images
  │      └─labels
  └─val
      ├─fog
      │  ├─images
      │  └─labels
      ├─night
      │  ├─images
      │  └─labels
      ├─rain
      │  ├─images
      │  └─labels
      └─snow
          ├─images
          └─labels

• You'll have to download the rgb_anon_trainvaltest.zip and gt_detection_trainval.zip in ACDC website if I had remember it correctly.

• Secondly, this framework requires all domains to be annotated, meaning both train/{source domain} and train/{target domain} must have annotations. And of course, all valid/* datasets must include object detection labels.

  • The images from different domains do not need to be paired, so the domain projection process is unsupervised.
  • Theoretically, you can remove the object detection loss for the target domain data, allowing the target domain to only involve unsupervised feature domain adaptation. This would fulfill the condition of "training without labeled target domain data."
    • If you do want to try this, the domain project should become global (since you don't have RoI info on target domain), so you'll probably have to modify the network architecture and dataloader.
    • I have tried this approach before, but the results were very unstable, likely due to the unsupervised nature of GANs. This instability led to significantly longer experiment periods (I only had 2 GPUs at the time), so I decided to label the data instead lol.

[dusongjie]

Thank you very much for your detailed response. I appreciate the clarification on the dataset structure and the requirements for the experiment. If I want to experiment on other datasets, they should follow the same structure, and both the source and target domain training images need labels, while the validation images of the target domain should also have labels.I appreciate your insights and the time you've taken to explain this. I will proceed with the labeled data approach as recommended and explore the potential modifications for unlabeled target domain data in future experiments.Thank you once again for your guidance and support. Best regards