Causal Debiasing for Visual Commonsense Reasoning

Jiayi Zou, Gengyun Jia, Bing-Kun Bao
10/23/2025
cs.CVcs.MM

No official code found. Best implementation: trea1262/VCR-OOD (1 star, MIT, last commit 2024-12-25).

Abstract

Visual Commonsense Reasoning (VCR) refers to answering questions and providing explanations based on images. While existing methods achieve high prediction accuracy, they often overlook bias in datasets and lack debiasing strategies. In this paper, our analysis reveals co-occurrence and statistical biases in both textual and visual data. We introduce the VCR-OOD datasets, comprising VCR-OOD-QA and VCR-OOD-VA subsets, which are designed to evaluate the generalization capabilities of models across two modalities. Furthermore, we analyze the causal graphs and prediction shortcuts in VCR and adopt a backdoor adjustment method to remove bias. Specifically, we create a dictionary based on the set of correct answers to eliminate prediction shortcuts. Experiments demonstrate the effectiveness of our debiasing method across different datasets.

View on arXivView PDF

Code Implementations(1)

trea1262/VCR-OODCommunity70%
MIT

The code is for Causal Debiasing for Visual Commonsense Reasoning.

10PythonDec 25, 20241 years agoMIT

Similar Papers

Integrating Object Interaction Self-Attention and GAN-Based Debiasing for Visual Question Answering

87% match1 repo112025

Common-Sense Bias Modeling for Classification Tasks

85% match2024

A Causal Adjustment Module for Debiasing Scene Graph Generation

84% match2025

A Comprehensive Survey on Visual Question Answering Datasets and Algorithms

83% match2024

Did Models Sufficient Learn? Attribution-Guided Training via Subset-Selected Counterfactual Augmentation

83% match2025

Building Reasonable Inference for Vision-Language Models in Blind Image Quality Assessment

83% match2025

Cite this paper

@article{zou2025causal,
  title  = {Causal Debiasing for Visual Commonsense Reasoning},
  author = {Jiayi Zou and Gengyun Jia and Bing-Kun Bao},
  year   = {2025},
  eprint = {2510.20281},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2510.20281v1}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2510.20281.svg)](https://codeofpaper.com/paper/2510.20281)

Discussion