submodules/loftup/README.md

LoftUp: A Coordinate-Based Feature Upsampler for Vision Foundation Models

ICCV2025 (oral)

[Arxiv] [Project Page]

Haiwen Huang, Anpei Chen, Volodymyr Havrylov, Andreas Geiger, Dan Zhang

TL;DR: LoftUp achieves the strongest feature upsampling performance at a comparable speed to bilinear upsampling.

Contents

• Install
• Inference with pretrained upsamplers
• Evaluation on downstream tasks
• Training LoftUp upsamplers
• Citation

Install

In general, LoftUp can run with most recent pytorch environments. We encourage the users to try out LoftUp in their exisitng environment first.

We also provide two yaml file for installation. To use them, simply run:

conda env create -f environment_cuda11.yaml

or

conda env create -f environment.yaml

Inference with pretrained upsamplers

All pre-trained upsamplers are available on 🤗 here: https://huggingface.co/models?search=loftup.

We provide example code for using LoftUp in example_usage.py. Currently we provide:

Backbone Name	Featurizer Class	HF hub	Torch Hub Repo	Torch Hub Name
DINOv2 S/14	dinov2	haiwen/loftup-dinov2s	andrehuang/loftup	loftup_dinov2s
DINOv2 S/14 + Reg	dinov2s_reg	haiwen/loftup-dinov2s_reg	andrehuang/loftup	loftup_dinov2s_reg
DINOv2 B/14	dinov2b	haiwen/loftup-dinov2b	andrehuang/loftup	loftup_dinov2b
DINOv2 B/14 + Reg	dinov2b_reg	haiwen/loftup-dinov2b_reg	andrehuang/loftup	loftup_dinov2b_reg
CLIP ViT B/16	clip	haiwen/loftup-clip	andrehuang/loftup	loftup_clip
SigLIP ViT B/16	siglip	haiwen/loftup-siglip	andrehuang/loftup	loftup_siglip
SigLIP2 ViT B/16	siglip2	haiwen/loftup-siglip2	andrehuang/loftup	loftup_siglip2

To use torch hub checkpoints, simply run

upsampler = torch.hub.load('andrehuang/loftup', model_torch_hub_name, pretrained=True)

For example, upsampler = torch.hub.load('andrehuang/loftup', loftup_dinov2s, pretrained=True).

The upsampler class is defined at UpsamplerwithChannelNorm.

Evaluation on Downstream Tasks

Dataset Preparation

See Preparing Datasets for Evaluation.

Semantic Segmentation

For semantic segmentation, our implementation is adapted from FeatUp. You can use eval_seg.py by running:

python eval_seg.py  ++upsampler_path=/path/to/your/upsampler

You can also configure other hyper-parameters such as output_dir and dataset directory. The config file is configs/eval_seg.yaml.

Video Object Segmentation

For video object segmentation on DAVIS, our code is modified from the implementation in LiFT. Specifically, we first extract segmentaiton results by running:

    python eval_davis.py --dataroot your_davis_data_dir --model_type "dinov2" --output_dir your_output_dir --imsize 224 --upsampler_path=your_upsampler_path

Then run the following to get evaluation results:

python davis2017-evaluation/evaluation_method.py --davis_path /your_davis_data_dir --task semi-supervised --results_path your_output_dir/davis_vidseg_224 --imsize 224

Others

For interactive segmentation, please check out iSegProbe.

For open-vocabulary segmentation, please check out ProxyCLIP.

For depth and normal estimation, please check out Probe3D.

Training LoftUp

We will provide training code soon.

Citation

If you find our work helpful, please cite:

@misc{huang2025loftuplearningcoordinatebasedfeature,
      title={LoftUp: Learning a Coordinate-Based Feature Upsampler for Vision Foundation Models}, 
      author={Haiwen Huang and Anpei Chen and Volodymyr Havrylov and Andreas Geiger and Dan Zhang},
      year={2025},
      eprint={2504.14032},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.14032}, 
}