Causal Inference with Sample Balancing for Out-of-Distribution Detection in Visual Classification

Image classification algorithms are commonly based on the Independent and Identically Distribution (IID) assumption, but in practice, the Out-Of-Distribution (OOD) problem is widely existing, i.e., the contexts of images in the model predicting are usually unseen during training. In this case, existing models trained under the IID assumption are limiting generalization. Causal inference is an important method to enhance the out-of-distribution generalization of models by partitioning various contexts from data and leading models to learn context-invariant predictions in different situations. However, existing methods mostly have imbalance problems due to the lack of constraints when partitioning data, which weakens the improvement of generalization. Therefore, we propose a Balanced Partition Causal Inference (BP-Causal) method, which automatically generates fine-grained balanced data partitions in an unsupervised manner, thereby enhancing the generalization ability of models in different contexts. Experiments on the OOD datasets NICO and NICO++ demonstrate that BP-Causal achieves stable predictions on OOD data, and we also find that models using BP-Causal focus more accurately on the foreground of images compared with the existing causal inference method, which effectively improves the generalization ability.