Improving Open-World Class-Agnostic Object Detectors via Feature Distillation with Student-Aware Adaptation

Open-world class-agnostic object detection aims to localize all objects in images regardless of whether their categories are known during training. Most existing studies focus on unannotated objects in training images, neglecting objects that are absent from the training set, namely, unseen-unknown objects. Although large-scale foundation models can potentially detect such unseen-unknown objects, their high computational demands limit their feasibility on real-world applications. Feature distillation can enable small yet effective detectors; however, existing methods struggle to transfer feature representations between detector pairs with differing architectures. To address these limitations, we propose a novel feature distillation approach that transfers feature representations from a foundation model to a small structurally different student detector. The proposed method effectively transfers these feature representations by explicitly adapting them to a student detector. Experimental results demonstrated that our approach significantly improves the detection performance of unseen-unknown objects across various detector types, including two-stage, one-stage, and anchor-free detectors.