Abstract:Class incremental learning aims to solve a problem that arises when continuously adding unseen class instances to an existing model This approach has been extensively studied in the context of image classification; however its applicability to object detection is not well established yet. Existing frameworks using replay methods mainly collect replay data without considering the model being trained and tend to rely on randomness or the number of labels of each sample. Also, despite the effectiveness of the replay, it was not yet optimized for the object detection task. In this paper, we introduce an effective buffer training strategy (eBTS) that creates the optimized replay buffer on object detection. Our approach incorporates guarantee minimum and hierarchical sampling to establish the buffer customized to the trained model. %These methods can facilitate effective retrieval of prior knowledge. Furthermore, we use the circular experience replay training to optimally utilize the accumulated buffer data. Experiments on the MS COCO dataset demonstrate that our eBTS achieves state-of-the-art performance compared to the existing replay schemes.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to effectively manage the replay buffer in Class Incremental Object Detection (CIOD) to reduce the phenomenon of catastrophic forgetting. Specifically, when the model is continuously adding new class instances, traditional replay methods (such as random replay) often cannot well preserve the knowledge of previous tasks, resulting in a decline in the model's detection performance for old classes. Therefore, the paper proposes an Effective Buffer Training Strategy (eBTS), aiming to improve the overall detection performance of the model on both new and old classes by optimizing the content of the replay buffer. ### Main Contributions 1. **Proposed an effective buffer management strategy**: This strategy selects samples based on two criteria, namely a high number of unique labels and a low loss value, rather than relying on a single metric (such as the number of labels, randomness, etc.). Experiments have verified that this method can effectively reflect the trend of the training model. 2. **Introduced the Circular Experience Replay (CER) training method**: This method overcomes the data imbalance problem caused by the limited capacity of the replay buffer by dealing with the asymmetry between current - task and previous - task data, thereby enhancing the performance of previous detections. ### Method Overview 1. **Guarantee Minimum Process**: Ensure that each class has a minimum number of samples in the buffer, reflecting the class distribution of the previous dataset, especially for rare classes. 2. **Hierarchical Sampling Strategy**: Optimize the buffer configuration to make it more representative by preferentially selecting samples with a high number of unique labels and a low loss value. 3. **Circular Experience Replay Training**: Combine the original Experience Replay (ER) training method to improve the utilization efficiency of previous knowledge by recycling the data in the buffer. ### Experimental Results The paper conducted experiments on the COCO dataset, and the results showed that eBTS achieved the best mAP (mean Average Precision)指标 in multiple scenarios, especially in settings where a large number of classes need to be retained. In addition, the ablation study further proved the effectiveness of CER in improving the model performance. ### Conclusion The method eBTS proposed in the paper significantly improves the model performance in class - incremental object detection tasks and reduces the phenomenon of catastrophic forgetting by optimizing the replay buffer management and training strategy. Future work will explore the combination with other strategies to deal with the forgetting problem more effectively.

Class-Wise Buffer Management for Incremental Object Detection: An Effective Buffer Training Strategy

One-Shot Replay: Boosting Incremental Object Detection via Retrospecting One Object

IncDet: in Defense of Elastic Weight Consolidation for Incremental Object Detection

TEAL: New Selection Strategy for Small Buffers in Experience Replay Class Incremental Learning

A Class-Incremental Detection Method of Remote Sensing Images Based on Selective Distillation

Bridge Past and Future: Overcoming Information Asymmetry in Incremental Object Detection

Bridging Non Co-occurrence with Unlabeled In-the-wild Data for Incremental Object Detection

Incremental Learning of Object Detectors without Catastrophic Forgetting

Incremental Object Detection with CLIP

Augmented Box Replay: Overcoming Foreground Shift for Incremental Object Detection

Static-Dynamic Co-Teaching for Class-Incremental 3D Object Detection

Incremental Detection of Remote Sensing Objects with Feature Pyramid and Knowledge Distillation

Continual Recognition with Adaptive Memory Update.

Dynamic Replay Training for Class-Incremental Learning

Enhancing class-incremental object detection in remote sensing through instance-aware distillation

SDDGR: Stable Diffusion-based Deep Generative Replay for Class Incremental Object Detection

IOR: Inversed Objects Replay for Incremental Object Detection

Towards Class-incremental Object Detection with Nearest Mean of Exemplars

Towards Non Co-occurrence Incremental Object Detection with Unlabeled In-the-Wild Data

RD-IOD: Two-Level Residual-Distillation-Based Triple-Network for Incremental Object Detection

CIOD: an intelligent class-incremental object detection system with nearest mean of exemplars