IOR: Inversed Objects Replay for Incremental Object Detection

TL;DR

This work tackles catastrophic forgetting in Incremental Object Detection when old-class objects are absent in incremental data (non co-occurrence). It introduces Inversed Objects Replay (IOR), which generates old-class samples by inverting the original detector, eliminating the need for extra generative models. IOR also employs augmented replay to reuse generated objects and a high-value distillation strategy that focuses on old-class signals amidst background, formalized as ${\mathcal{L}_{total}} = {\mathcal{L}_{detect}} + {\mathcal{L}_{dist-feat}} + {\mathcal{L}_{dist-logit}}$, with ROIAlign-based feature distillation on replayed boxes. Experiments on MS COCO 2017 show state-of-the-art performance in non co-occurrence with no extra model training and competitive results in co-occurrence, highlighting the approach's practical efficiency for real-world IOD settings.

Abstract

Existing Incremental Object Detection (IOD) methods partially alleviate catastrophic forgetting when incrementally detecting new objects in real-world scenarios. However, many of these methods rely on the assumption that unlabeled old-class objects may co-occur with labeled new-class objects in the incremental data. When unlabeled old-class objects are absent, the performance of existing methods tends to degrade. The absence can be mitigated by generating old-class samples, but it incurs high costs. This paper argues that previous generation-based IOD suffers from redundancy, both in the use of generative models, which require additional training and storage, and in the overproduction of generated samples, many of which do not contribute significantly to performance improvements. To eliminate the redundancy, we propose Inversed Objects Replay (IOR). Specifically, we generate old-class samples by inversing the original detectors, thus eliminating the necessity of training and storing additional generative models. We propose augmented replay to reuse the objects in generated samples, reducing redundant generations. Moreover, we propose high-value knowledge distillation focusing on the positions of old-class objects overwhelmed by the background, which transfers the knowledge to the incremental detector. Extensive experiments conducted on MS COCO 2017 demonstrate that our method can efficiently improve detection performance in IOD scenarios with the absence of old-class objects. The code is available at https://github.com/JiaJia075/IOR.

Figures (4)

• Figure 1: Compared with other methods (SDDGRkim2024sddgr and PseudoRMyang2023pseudo), IOR generates old-class objects without extra generators, efficiently addressing the absence of old-class objects in non co-occurrence scenarios.
• Figure 2: Visualization results for generating 1 vs. 30 objects.
• Figure 3: Framework of Inversed Objects Replay for incremental object detection.
• Figure 4: Comparison of visualization results with and without $\mathcal{R}(I)$ for generated object.