Robust Unknown Object Detection in Dynamic Environments through Dual-Granularity Reconstruction Error Modeling

Open World Object Detection (OWOD) aims to detect both known and unknown objects in dynamic environments, where unknown instances lack ground-truth supervision during training. Existing methods typically rely on supervision from known categories, leading models to overconfidently classify visually similar unknowns as known classes, and dissimilar ones as background. This known-class prior bias severely hinders the detection of truly novel objects. To address this challenge, we propose a robust unknown object detection method based on dual-granularity reconstruction error modeling. At the fine-grained level, we propose Fine-grained Masked Reconstruction (FMR), which randomly masks feature regions to guide reconstruction toward semantic structures, thereby improving foreground–background discrimination. At the coarse-grained level, we propose Adaptive Region-based Error Aggregation (AREA), which aggregates reconstruction errors over object proposals to enhance the model’s sensitivity to ambiguous semantic boundaries while suppressing local outliers. Furthermore, we perform decoupled probabilistic modeling of foreground and background reconstruction errors, enabling soft estimation of unknown object likelihoods without supervision. Extensive experiments on standard OWOD benchmarks demonstrate that our method consistently outperforms state-of-the-art approaches, achieving a +20.6 improvement in unknown object recall (U-Recall) while maintaining strong performance on known classes.