Abstract

Recently, object detection models have witnessed notable performanceimprovements, particularly with transformer-based models. However, new objectsfrequently appear in the real world, requiring detection models to continuallylearn without suffering from catastrophic forgetting. Although IncrementalObject Detection (IOD) has emerged to address this challenge, these existingmodels are still not practical due to their limited performance and prolongedinference time. In this paper, we introduce a novel framework for IOD, calledHier-DETR: Hierarchical Neural Collapse Detection Transformer, ensuring bothefficiency and competitive performance by leveraging Neural Collapse forimbalance dataset and Hierarchical relation of classes' labels.

 

Quick Read (beta)