Mitigating Context Bias via Foreground-Background Separation and Pooling: A Causal Analysis and Robust Evaluation

ICLR 2026 Conference Submission18577 Authors

Keywords: pooling, context bias, object detection

Abstract:

Context bias refers to the association between the foreground objects and background during the object detection training process. Various methods have been proposed to minimize the context bias when applying the trained model to an unseen domain, known as domain adaptation for object detection (DAOD). But a principled approach to understand why the context bias occurs and how to remove it has been missing.
In this work, we provide a causal view of the context bias in the network architecture as the possible source of this bias. We present an analytical framework that utilizes an explicit foreground mask during feature aggregation with the proposed pooling operation to separate foreground and background, which leads the trained model to detect objects in a more robust manner under different domains. We use the ground truth masks and also masks generated using Segment Anything Model (SAM) to showcase the performance of the different state-of-the-art network model architectures such as ALDI++, ResNet, EfficientNet and Vision Transformer. We also provide a benchmark designed to create an ultimate test for DAOD, using foregrounds in the presence of absolute random backgrounds, to statistically analyze the robustness of the intended trained models using 95\% confidence. Through these experiments, our goal is to provide a principled approach for minimizing context bias under domain shift for object detection.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 18577