Temel Kavramlar
Inversed Objects Replay (IOR) efficiently generates old-class object samples to mitigate catastrophic forgetting in incremental object detection without co-occurrence of old and new class objects.
Özet
The paper proposes Inversed Objects Replay (IOR) to efficiently address the performance degradation of incremental object detection (IOD) methods in non co-occurrence scenarios, where images in the incremental dataset lack old-class objects.
Key highlights:
- IOR generates old-class object samples by inversing the original detector, eliminating the need for training and storing additional generative models required by previous generation-based IOD methods.
- IOR employs augmented replay to reuse the generated objects, reducing the requirement for generating massive samples.
- IOR introduces high-value distillation to focus on distilling the outputs relevant to old-class objects, mitigating the interference from the background.
Extensive experiments on MS COCO 2017 dataset demonstrate that IOR can efficiently improve detection performance in IOD scenarios with the absence of old-class objects, outperforming state-of-the-art distillation-based and generation-based IOD methods.
İstatistikler
The paper reports the average precision (AP) results on the MS COCO 2017 dataset for one-step and multi-step IOD settings under both co-occurrence and non co-occurrence scenarios.
Alıntılar
"To mitigate catastrophic forgetting for non co-occurrence IOD with low costs, we propose the Inversed Objects Replay (IOR)."
"We argue that the extra cost stems from the redundancy of generative models and sample generations. We exclude redundant generative models by inversing the original detector, eliminating the necessity of training or saving generative models."
"To effectively utilize the generated objects, we distill incremental data with replayed objects. However, the generated objects are overwhelmed by the background, leading to ineffective distillation. Therefore, we propose high-value knowledge distillation, focusing on distilling outputs relevant to old-class objects."