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

Forum topic · 小凯 · 2026-03-31

Summary

This arXiv paper (2503.23724) by Md Ashiqur Rahman, Lim Jun Hao, and Jeremiah Jiang proposes a general framework for building soft equivariant computer vision models. Instead of enforcing strict equivariance—which real-world data rarely satisfies—the method projects model weights into a designed subspace, allowing the degree of equivariance to be tuned. The approach works on any pre-trained architecture and comes with theoretical bounds on the induced equivariance error. The authors validate it on multiple pre-trained backbones, including ViT and ResNet, across image classification, semantic segmentation, and human trajectory prediction tasks. Notably, the method improves performance while reducing equivariance error on the competitive ImageNet benchmark, making it a practical and theoretically grounded way to control equivariance in vision models.

Paper Overview

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

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.

The authors propose a general framework for constructing soft equivariant models by projecting the model weights into a designed subspace. This method:

Applies to any pre-trained architecture
Provides theoretical bounds on the induced equivariance error
Enables tuning the degree of equivariance rather than enforcing strict equivariance

Empirical Results

The method is validated on multiple pre-trained backbones, including ViT and ResNet, across three task types:

Image classification
Semantic segmentation
Human trajectory prediction

Notably, the approach improves performance while simultaneously reducing equivariance error on the highly competitive ImageNet benchmark.

Source

Paper: Tunable Soft Equivariance with Guarantees

*Auto-collected on 2026-03-31*