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.

Class Incremental Robotic Pick-and-Place via Incremental Few-Shot Object Detection

Incremental Few-Shot Object Detection for Robotics

A Cognition-Driven Framework for Few-Shot Class-Incremental Learning

Expandable-RCNN: toward high-efficiency incremental few-shot object detection

Few Shot Class Incremental Learning via Efficient Prototype Replay and Calibration

Towards Generalized and Incremental Few-Shot Object Detection

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

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

Robotic pick-and-place of novel objects in clutter with multi-affordance grasping and cross-domain image matching

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

A Robotic Semantic Grasping Method for Pick-and-place Tasks

UIFormer: A Unified Transformer-based Framework for Incremental Few-Shot Object Detection and Instance Segmentation

InfRS: Incremental Few-Shot Object Detection in Remote Sensing Images

Learning Few-Shot Object Placement with Intra-Category Transfer

Revisiting Class-Incremental Object Detection: an Efficient Approach Via Intrinsic Characteristics Alignment and Task Decoupling

Incremental Object Detection with CLIP

An Incremental Knowledge Learning Framework for Continuous Defect Detection

Active Class Selection for Few-Shot Class-Incremental Learning

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

A Forward and Backward Compatible Framework for Few-shot Class-incremental Pill Recognition

Class-Incremental Few-Shot Object Detection