Abstract:With the rapid development of remote sensing technology and the growing demand for applications, the classical deep learning-based object detection model is bottlenecked in processing incremental data, especially in the increasing classes of detected objects. It requires models to sequentially learn new classes of objects based on the current model, while preserving old categories-related knowledge. Existing class-incremental detection methods achieve this goal mainly by constraining the optimization trajectory in the feature of output space. However, these works neglect the case where the previously learned background is a new category to learn, resulting in performance degradation in the new category because of the conflict between remaining the background-related knowledge or updating the background-related knowledge. This paper proposes a novel class-incremental detection method incorporated with the teacher-student architecture and the selective distillation (SDCID) strategy. Specifically, it is the asymmetry architecture, i.e., the teacher network temporarily stores historical knowledge of previously learned objects, and the student network integrates historical knowledge from the teacher network with the newly learned object-related knowledge, respectively. This asymmetry architecture reveals the significance of the distinct representation of history knowledge and new knowledge in incremental detection. Furthermore, SDCID selectively masks the shared subobject of new images to learn and previously learned background, while learning new categories of images and then transfers the classification results of the student model to the background class following the judgment model of the teacher model. This manner avoids interferences between the background category-related knowledge from a teacher model and the learning of other new classes of objects. In addition, we proposed a new incremental learning evaluation metric, C-SP, to comprehensively evaluate the incremental learning stability and plasticity performance. We verified the proposed method on two object detection datasets of remote sensing images, i.e., DIOR and DOTA. The experience results in accuracy and C-SP suggest that the proposed method surpasses existing class-incremental detection methods. We further analyzed the influence of the mask component in our method and the hyper-parameters sensitive to our method.

Context-aware Feature Reconstruction for Class-Incremental Anomaly Detection and Localization

A Diffusion-Based Framework for Multi-Class Anomaly Detection

Anomaly Detection by Adapting a pre-trained Vision Language Model

Contextual Affinity Distillation for Image Anomaly Detection

Pull & Push: Leveraging Differential Knowledge Distillation for Efficient Unsupervised Anomaly Detection and Localization

PairCFR: Enhancing Model Training on Paired Counterfactually Augmented Data through Contrastive Learning

Learning Prompt-Enhanced Context Features for Weakly-Supervised Video Anomaly Detection

Towards Continual Adaptation in Industrial Anomaly Detection

Revitalizing Reconstruction Models for Multi-class Anomaly Detection via Class-Aware Contrastive Learning

AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios

Time-series Anomaly Detection via Contextual Discriminative Contrastive Learning

DA-CIL: Towards Domain Adaptive Class-Incremental 3D Object Detection

Continuous Memory Representation for Anomaly Detection

Enhancing Multi-Class Anomaly Detection via Diffusion Refinement with Dual Conditioning

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

Class-Incremental Learning Based on Anomaly Detection

Enhanced multi-scale features mutual mapping fusion based on reverse knowledge distillation for industrial anomaly detection and localization

Learning to Classify With Incremental New Class

Customizing Visual-Language Foundation Models for Multi-modal Anomaly Detection and Reasoning

Advancing Pre-trained Teacher: Towards Robust Feature Discrepancy for Anomaly Detection

CLIP3D-AD: Extending CLIP for 3D Few-Shot Anomaly Detection with Multi-View Images Generation