Incremental Object Detection System For Open World

In recent years,deep learning-based object detection models have made impressive achievements.However,current object detection algorithms cannot be directly applied to open environment due to the following two challenges that open environment bring to the models:(1)open environment is complex and variable,with many unknown category samples in the scenes.Ordinary object detection models are difficult to adapt to complex open environment and require frequent incremental learning retraining.This leads to extremely high training costs,which affects the practical application of the model on the ground.Therefore,low-cost incremental object detection models for open environment need to be designed.(2)The data distribution in the open environment is uneven,and the amount of sample data in some categories is scarce.The existing models have weak feature extraction ability for few-shot data,and training on few-shot data will produce serious overfitting problems,resulting in poor detection performance.Therefore,a few-shot incremental object detection model for open environment needs to be designed.To address the above problems and challenges,this paper carries out the following work:(1)To address the problem of high cost of retraining incremental object detection models in open environment,this paper designs and implements a low-cost incremental object detection model that including an adaptive unknown category sample discovery method and a cross-modal semantic classification method.The model performance is not affected while reducing the number of parameters and dataset production cost to achieve low-cost incremental object detection training.(2)To address the problem of poor model performance in open environment few-shot training scenarios,this paper designs and implements a few-shot incremental object detection algorithm that including a gradient control strategy and a data extrapolation-based data augmentation method.The gradient control strategy can balance the weight of each sub-task gradient on the model performance and effectively alleviate the overfitting problem of model training in few-shot scenarios.The data extrapolation-based data augmentation method performs data augmentation for data with extremely scarce sample size,which effectively improves the detection performance of the model.This algorithm performs well on the fewshot incremental object detection task and achieves SOTA performance on CNN based-architecture for this type of task.(3)To simplify model development,model deployment,etc.In this paper,we design and implement an incremental object detection system for open environment.The system encapsulates the development process of related algorithms and provides data management,task management,model management and other related services for system users,which reduces the workload of model developers.