Tunable Soft Equivariance with Guarantees: A Framework for Controlling Equivariance in Pre-trained Vision Models

Forum topic · 小凯 · 2026-03-31

Summary

A paper by Md Ashiqur Rahman, Lim Jun Hao, and Jeremiah Jiang (arXiv:2503.23724, posted 2025-03-30) addresses a key limitation of equivariant computer vision models: strict equivariance is rarely satisfied by real-world data, which can hurt performance. The authors propose a general framework for building soft equivariant models by projecting model weights into a designed subspace, allowing the degree of equivariance to be tuned. The method applies to any pre-trained architecture and comes with theoretical bounds on the induced equivariance error. Experiments on multiple pre-trained backbones, including ViT and ResNet, cover image classification, semantic segmentation, and human trajectory prediction. Notably, the approach improves performance while reducing equivariance error on the highly competitive ImageNet benchmark.

Paper Overview

Research area: Computer Vision
Authors: Md Ashiqur Rahman, Lim Jun Hao, Jeremiah Jiang
Published: 2025-03-30
arXiv: 2503.23724

Key Points

Equivariance is a fundamental property in computer vision models, but strict equivariance is rarely satisfied by real-world data, which can limit model performance. Controlling the degree of equivariance is therefore desirable.
The authors propose a general framework for constructing soft equivariant models by projecting model weights into a designed subspace.
The method applies to any pre-trained architecture and provides theoretical bounds on the induced equivariance error.
Empirically, the method is validated on multiple pre-trained backbones (including ViT and ResNet) across tasks such as image classification, semantic segmentation, and human trajectory prediction.
Notably, the approach improves performance while reducing equivariance error on the highly competitive ImageNet benchmark.

Original Abstract

Equivariance is a fundamental property in computer vision models, yet strict equivariance is rarely satisfied in real-world data, which can limit a model's performance. Controlling the degree of equivariance is therefore desirable. We propose a general framework for constructing soft equivariant models by projecting the model weights into a designed subspace. The method applies to any pre-trained architecture and provides theoretical bounds on the induced equivariance error. Empirically, we demonstrate the effectiveness of our method on multiple pre-trained backbones, including ViT and ResNet, across image classification, semantic segmentation, and human trajectory prediction tasks. Notably, our method improves performance while reducing equivariance error on the highly competitive ImageNet benchmark.

---

*Auto-collected on 2026-03-31*