Domain incremental learning for object detection

doi:10.1016/j.patcog.2025.111882

Return

Domain incremental learning for object detection

2025-06-05

PRE

DOI：10.1016/j.patcog.2025.111882

Original

Original request for help

Save

Abstract

En 中文

Incremental learning is an important research spot that can help the model continually adapt to new data without forgetting old knowledge. In the field of object detection, existing works mainly focus on new-class incremental learning. However, in many practical applications, the well pre-trained object detection systems often suffer from abrupt performance degradations when adapting to new data domain, i.e., the drastic changes in object size, class distribution and image context. To this end, we propose a new-instance incremental learning task for object detection, called Domain Incremental Learning with Limited Budgets (DILLB). DILLB adopts the conventional incremental setting to pre-train the detection network on the source dataset and incrementally fine-tune the network on the target domain. Its objective is to maximize the performance of the detector on the target dataset, while avoiding catastrophic forgetting on the source one. Meanwhile, to increase the challenge of DILLB, we also place strict restrictions on the access to label information of both source and target datasets, helping DILLB get closer to the practical applications. In terms of baselines, we not only reconstruct common approaches of incremental learning and domain adaption, but also propose a novel teacher–student based method for domain incremental object detection, which demonstrates obvious merits in improving performance and alleviating catastrophic forgetting. In this paper, we also propose a comprehensive experimental setup for DILLB to promote its research in object detection. Our source code is available at https://github.com/Disguiser15/DILLB .