baogege1995/FPS_H5

FPS-UDA Feature Banks

This repository hosts released dataset-level H5 feature banks for FPS-UDA (Feature-Space Planes Searcher for Unsupervised Domain Adaptation).

The feature banks are used by the FPS-UDA codebase:

• GitHub: https://github.com/baogegeJiang/FPS
• Documentation: https://baogegeJiang.github.io/FPS/
• Paper: https://ieeexplore.ieee.org/abstract/document/11568428

What Is Included

The files are stored under the banks/ directory.

Key	File	Backbone family
office31_resnet	banks/office31_resnet50.h5	torchvision ResNet
office31_vit	banks/office31_vit.h5	timm ViT
office31_siglip2	banks/office31_siglip2.h5	HuggingFace AutoModel / SigLIP 2
office_home_resnet	banks/office_home_resnet50.h5	torchvision ResNet
office_home_vit	banks/office_home_vit.h5	timm ViT
office_home_siglip2	banks/office_home_siglip2.h5	HuggingFace AutoModel / SigLIP 2
visda17_resnet	banks/visda17_resnet101.h5	torchvision ResNet
visda17_vit	banks/visda17_vit.h5	timm ViT
visda17_siglip2	banks/visda17_siglip2.h5	HuggingFace AutoModel / SigLIP 2

The released FPS-UDA library supports both classic vision backbones and newer large vision models. In particular, the SigLIP 2 banks are extracted with:

backbone:
  backend: hf_auto_vision
  name: google/siglip2-so400m-patch14-384