Image Background Serves as Good Proxy for Out-of-distribution Data

TL;DR

This work addresses the challenge of out-of-distribution detection by formulating a probabilistic framework grounded in Bayes' rule that factorizes $P(w_i|x)$ into an ID-classification component and an explicit OOD factor $P(x\in\mathcal{S}_{ID}|x)$. It then introduces SSOD, an end-to-end method that self-supervises OOD signals by sampling from ID image backgrounds, avoiding explicit OOD data or synthetic generation. The approach yields strong empirical gains across ImageNet and CIFAR-10 benchmarks, including hard OOD sets like ImageNet-O and OpenImage-O, demonstrating improved ID/OOD separation and robustness. Overall, the work provides a unified interpretation of existing OOD methods and a scalable, data-efficient mechanism for open-world detection with practical impact for real-world systems.

Abstract

Out-of-distribution (OOD) detection empowers the model trained on the closed image set to identify unknown data in the open world. Though many prior techniques have yielded considerable improvements in this research direction, two crucial obstacles still remain. Firstly, a unified perspective has yet to be presented to view the developed arts with individual designs, which is vital for providing insights into future work. Secondly, we expect sufficient natural OOD supervision to promote the generation of compact boundaries between the in-distribution (ID) and OOD data without collecting explicit OOD samples. To tackle these issues, we propose a general probabilistic framework to interpret many existing methods and an OOD-data-free model, namely \textbf{S}elf-supervised \textbf{S}ampling for \textbf{O}OD \textbf{D}etection (SSOD). SSOD efficiently exploits natural OOD signals from the ID data based on the local property of convolution. With these supervisions, it jointly optimizes the OOD detection and conventional ID classification in an end-to-end manner. Extensive experiments reveal that SSOD establishes competitive state-of-the-art performance on many large-scale benchmarks, outperforming the best previous method by a large margin, \eg, reporting \textbf{-6.28\%} FPR95 and \textbf{+0.77\%} AUROC on ImageNet, \textbf{-19.01\%} FPR95 and \textbf{+3.04\%} AUROC on CIFAR-10, and top-ranked performance on hard OOD datasets, \ie, ImageNet-O and OpenImage-O.

Figures (16)

• Figure 1: Feature visualization of the ID and OOD images. The green/orange dots are surrogate ID/OOD features generated by SSOD. The blue/gray dots are natural ID/OOD features of ImageNet/iNaturalist.
• Figure 2: Locality of ResNet-50.
• Figure 3: Locality of SSOD.
• Figure 4: Samplers of ResNet-50.
• Figure 5: Samplers of SSOD.