Hunting for the unknown: Open world object detection from a class-agnostic perspective.

Abstract

The existing Open World Object Detection models rely on pseudo-labeling to annotate unknown objects during training. However, this approach leads to an over-dependence on known objects, thereby weakening the model's capability to detect unknown objects. To tackle this issue, this paper presents a novel class-agnostic object detection model based on dynamic foreground perception and localization. The model leverages a dynamic foreground perception and localization algorithm that adeptly distinguishes foreground and background regions within images using dynamic detection heads. Additionally, by employing class-agnostic detection that does not rely on specific class information, the model mitigates excessive dependence on known categories and demonstrates improved performance in the detection of unknown objects. The key innovation of the model revolves around three main aspects: the refinement of spatial perception features, the disentanglement of attention features, and dynamic foreground perception and localization. Experimental findings across PASCAL VOC, COCO2017, LVISv1.0, and Objects365 datasets demonstrate that our model maintains high-level detection performance on known objects while surpassing most existing methods in the detection of unknown objects, exhibiting +11 points improvement in U-Recall performance. These results affirm the efficacy and superiority of the proposed detection method detailed in this paper.