LoftUp: Learning a Coordinate-Based Feature Upsampler for Vision Foundation Models

Haiwen Huang, Anpei Chen, Volodymyr Havrylov, Andreas Geiger, Dan Zhang
arXiv Code
Example 1
Example 2
Example 3

Abstract

Vision foundation models (VFMs) such as DINOv2 and CLIP have achieved impressive results on various downstream tasks, but their limited feature resolution hampers performance in applications requiring pixel-level understanding. Feature upsampling offers a promising direction to address this challenge.

In this work, we identify two critical factors for enhancing feature upsampling: the upsampler architecture and the training objective. For the upsampler architecture, we introduce a coordinate-based cross-attention transformer that integrates the high-resolution images with coordinates and low-resolution VFM features to generate sharp, high-quality features. For the training objective, we propose constructing high-resolution pseudo-groundtruth features by leveraging class-agnostic masks and self-distillation. Our approach effectively captures fine-grained details and adapts flexibly to various input and feature resolutions. Through experiments, we demonstrate that our approach significantly outperforms existing feature upsampling techniques across various downstream tasks.

LoftUp significantly enhances VFM performance on various downstream tasks

performance

How does it work?

We identified two key factors to the success of feature upsampling: upsampler architecture and training objective.

cars peace

For the upsampler architecture, we introduce a coordinate-based cross-attention transformer that integrates the high-resolution images with coordinates and low-resolution VFM features to generate sharp, high-quality features.

cars peace

For the training objective, we propose constructing high-resolution pseudo-groundtruth features by leveraging class-agnostic masks and self-distillation.

Results

Left: Feature Visualization. Middle: Depth Estimation (downstream). Right: Semantic Segmentation (downstream).

GIF 1
GIF 2
GIF 3
GIF 4
GIF 5
GIF 6

BibTeX

@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}, 
}