Research On Sampling Optimization For CNN-based Object Detectors

Object detection is a pivotal realm in the field of computer vision.It is widely applied in many hot realms,like auto-driving car,computer-aided diagnosis and traffic condition monitoring.In the process of object detection,many regions are sampled from the input image,constituting a large sample space.As the sparsity of foreground samples,background samples are dominant in the sample space,leading to the problem of imbalance between foreground and background in object detection and limiting the further improvement of detection accuracy.For region-based object detectors,they adopt region-proposing stage and subsequent sampling strategies which rely on that stage to drop backgrounds in the sampling space effectively,which boosts the detection accuracy of detectors.In the original developing stage of CNN(Convolutional Neural Network)-based object detectors,region-based detectors are in the leading position,as detection accuracy is the bottleneck of object detection in that stage.The application of region-based detectors,however,becomes really limited because the low detection speed and complex computation in region-based detectors cannot follow the treads of the steady improvement of detection accuracy and the strong demand of real-time detection.In contrast,unified pipeline detectors that abandon the use of region-proposing stage and lead to higher detection speed have received widespread attention in recent years.But unified pipeline detectors suffer from low detection accuracy for the huge imbalance between backgrounds and foregrounds in the sampling space that detectors must face with.Though many efforts try to propose sampling strategies for unified pipeline detectors and they do make some progresses,they are not general methods and they rely heavily on specific structures or loss functions,and the consequence is that many state-of-the-art object detectors like YOLO series have yet to benefit from those sampling strategies.This paper proposes LRM(Loss Rank Mining)method to fill the gap in general sampling methods for unified pipeline detectors.Firstly,we analyze the process of evolution of sampling optimization in CNN-based object detectors and put the sampling method at the end of unified pipeline detectors to maintain the their structures and advantages in detection rates.Secondly,we analyze the structure of unified pipeline detectors and exploit the output feature map which is a general structure used to present detection results in unified pipeline detectors to perform our sampling strategy,which makes LRM a general method in unified pipeline detectors.Finally,we design LRM as a training strategy,which introduces no additional computation in the process of inference,improving detection accuracy in the precondition of sacrificing no inference speed.By using LRM,region samples where detectors perform worse are sampled for training and those samples are more likely to be foreground,which makes the training of detectors concentrate more on foreground samples.Extensive experiments validate that LRM improve the detection accuracy of many state-of-the-art unified pipeline detectors on many datasets.This paper analyze the difference importance of classification and localization in diverse stages of training for the same input and the loss value distribution of region samples for different inputs.Also,we consider LRM's shortage of localization accuracy and fitting different input data and propose ALRM(Adaptive Loss Rank Mining)method by introducing the idea of attention and data-driven into LRM method.ALRM is a sampling optimization method and could improve detection accuracy via improving robustness in two aspects: firstly,in the two tasks of the same input image,we improve the localization accuracy through adjusting the attention that the model put on classification task or localization task;secondly,in the sampling of tons of region proposals from different input images,we sample region proposals according to the distribution of loss values instead of sampling a fixed number of region proposals,further improving the robustness of detectors.Also,ALRM avoid the use of hyper-parameters,making the process of training much easier.Extensive experiments validate that the detection accuracy of many state-ofthe-art unified-pipeline detectors improved by ALRM exceed that improved by existing works on several datasets.