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

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Existing Open World Object Detection models use pseudo-labeling, leading to over-reliance on known objects.
• This over-dependence hinders the models' ability to detect unknown objects effectively.

Purpose of the Study:

• To present a novel class-agnostic object detection model.
• To improve the detection of unknown objects by mitigating dependence on known categories.

Main Methods:

• Developed a class-agnostic object detection model utilizing dynamic foreground perception and localization.
• Employed dynamic detection heads for distinguishing foreground and background regions.
• Incorporated refinement of spatial perception features and disentanglement of attention features.

Main Results:

• The model maintains high detection performance on known objects.
• Significantly surpasses existing methods in detecting unknown objects, showing a +11 points improvement in U-Recall.
• Validated on PASCAL VOC, COCO2017, LVISv1.0, and Objects365 datasets.

Conclusions:

• The proposed dynamic foreground perception and localization method is effective for Open World Object Detection.
• The class-agnostic approach enhances the detection of unknown objects.
• The model demonstrates superior performance compared to existing methods for unknown object detection.