[2505.16985] Extremely Simple Multimodal Outlier Synthesis for Out-of-Distribution Detection and Segmentation

Computer Science > Computer Vision and Pattern Recognition arXiv:2505.16985 (cs) [Submitted on 22 May 2025 (v1), last revised 4 Mar 2026 (this version, v2)] Title:Extremely Simple Multimodal Outlier Synthesis for Out-of-Distribution Detection and Segmentation Authors:Moru Liu, Hao Dong, Jessica Kelly, Olga Fink, Mario Trapp View a PDF of the paper titled Extremely Simple Multimodal Outlier Synthesis for Out-of-Distribution Detection and Segmentation, by Moru Liu and 4 other authors View PDF HTML (experimental) Abstract:Out-of-distribution (OOD) detection and segmentation are crucial for deploying machine learning models in safety-critical applications such as autonomous driving and robot-assisted surgery. While prior research has primarily focused on unimodal image data, real-world applications are inherently multimodal, requiring the integration of multiple modalities for improved OOD detection. A key challenge is the lack of supervision signals from unknown data, leading to overconfident predictions on OOD samples. To address this challenge, we propose Feature Mixing, an extremely simple and fast method for multimodal outlier synthesis with theoretical support, which can be further optimized to help the model better distinguish between in-distribution (ID) and OOD data. Feature Mixing is modality-agnostic and applicable to various modality combinations. Additionally, we introduce CARLA-OOD, a novel multimodal dataset for OOD segmentation, featuring synthetic OOD objects ...