FeatUp: A Model-Agnostic Framework for Features at Any Resolution

ICLR 2024

Stephanie Fu*, Mark Hamilton*, Laura Brandt, Axel Feldman, Zhoutong Zhang, William T. Freeman *Equal Contribution.

TL;DR:FeatUp improves the spatial resolution of any model's features by 16-32x without changing their semantics.

teaser.2.mp4

Contents

• Install
• Using Pretrained Upsamplers
• Fitting an Implicit Upsampler
• Coming Soon
• Citation
• Contact

Install

Pip

For those just looking to quickly use the FeatUp APIs install via:

pip install git+https://github.com/mhamilton723/FeatUp

Local Development

To install FeatUp for local development and to get access to the sample images install using the following:

git clone https://github.com/mhamilton723/FeatUp.git
cd FeatUp
pip install -e .

Using Pretrained Upsamplers

To see examples of pretrained model usage please see our Collab notebook. We currently supply the following pretrained versions of FeatUp's JBU upsampler:

Model Name	Checkpoint	Checkpoint (No LayerNorm)	Torch Hub Repository	Torch Hub Name
DINO	Download	Download	mhamilton723/FeatUp	dino16
DINO v2	Download	Download	mhamilton723/FeatUp	dinov2
CLIP	Download	Download	mhamilton723/FeatUp	clip
MaskCLIP	n/a	Download	mhamilton723/FeatUp	maskclip
ViT	Download	Download	mhamilton723/FeatUp	vit
ResNet50	Download	Download	mhamilton723/FeatUp	resnet50

For example, to load the FeatUp JBU upsampler for the DINO backbone without an additional LayerNorm on the spatial features:

upsampler = torch.hub.load("mhamilton723/FeatUp", 'dino16', use_norm=False)

To load upsamplers trained on backbones with additional LayerNorm operations which makes training and transfer learning a bit more stable:

upsampler = torch.hub.load("mhamilton723/FeatUp", 'dino16')

Fitting an Implicit Upsampler to an Image

To train an implicit upsampler for a given image and backbone first clone the repository and install it for local development. Then run

cd featup
python train_implicit_upsampler.py

Parameters for this training operation can be found in the implicit_upsampler config file.

Local Gradio Demo

To run our HuggingFace Spaces hosted FeatUp demo locally first install FeatUp for local development. Then run:

python gradio_app.py

Wait a few seconds for the demo to spin up, then navigate to http://localhost:7860/ to view the demo.

Coming Soon:

• Training your own FeatUp joint bilateral upsampler
• Simple API for Implicit FeatUp training

Citation

@inproceedings{
    fu2024featup,
    title={FeatUp: A Model-Agnostic Framework for Features at Any Resolution},
    author={Stephanie Fu and Mark Hamilton and Laura E. Brandt and Axel Feldmann and Zhoutong Zhang and William T. Freeman},
    booktitle={The Twelfth International Conference on Learning Representations},
    year={2024},
    url={https://openreview.net/forum?id=GkJiNn2QDF}
}

Contact

For feedback, questions, or press inquiries please contact Stephanie Fu and Mark Hamilton