Decoupled Visual Causality for Robust Detection

Abstract

The existing empirical risk minimization algorithms learn the association between inputs and labels, and face substantial difficulties when apply to different distributions because of various confounders. Causal intervention becomes a solid solution to this issue by analyzing the visual causality, instead, those approaches fail at disentangling the confounders and mediators within the causality, and bring negative effects to the prediction. In this paper, we propose a disentangled visual causal model to eliminate the effects of confounders while reserving the corresponding mediators. Specifically, confounders are considered as different objects on the image, while mediators are formulated as some critical components of the targets that contribute to a distinctive identification. Extensive experiments on coco datasets have demonstrated the superiority of our model over other state-of-the-art baselines.