Multi-stage Based Object Detection Post Processing Method In Overlapping Scene

With the continuous upgrading of computer hardware and the rise of deep learning,remarkable achievements have been made in the field of computer vision.Object detection algorithm is one of the most important research area in the field of computer vision.At present most of the object detection algorithms are three-stage process,the first stage is to generate region proposals,and then predict the score for the candidate box,and finally post-processing algorithm is used to select the best candidate box set as the output.The candidate box generation stage will produce a set of candidate boxes with a large number of redundancies for each target object.The goal of the post-processing algorithm is used to remove redundant candidates on the target object and keep one best matching box for each target object.Almost all current object detectors use non-maximum suppression algorithms as post-processing algorithm.This algorithm first select the candidate box with the highest confidence and then suppress all candidate boxes which proportion of overlap with the most confident box greater than a certain threshold,through this iterative process non-maximum suppression achieve post-processing algorithm's goal.In complex scenes where the objects are densely overlapped,even if the object detection algorithm can generate sufficiently good region proposals,the non-maximum suppression algorithm simply uses the overlap threshold to suppress or retain the candidate box could not perform well on precision and recall simultaneously.How to improve the object detection postprocessing algorithm for the improvement of the target detection algorithm in overlapping scenarios is extremely important.In this paper,we propose a multi-stage post-processing algorithm based on the characteristics of the candidate box generated by the object detector in overlapping scenes and the defects of the non-maximum suppression algorithm.The algorithm completes the post-processing process through four stages: regional division,target quantity prediction,selection of candidate boxes in dense area and mergence of results.The division process divides the picture into several sub-regions with different target densities by clustering method.By counting the distribution features and high-order features of the candidate boxes in the dense region,we counld predict the target amount to assist the post-processing process.Based on the forecasting target numbers,our algorithm will divide candidate boxes to a fixed number of clusters,while the number can also be used as a constraint to limit the determinant point process algorithm,and then completed the selection of the candidate boxes inside the dense area.The output of the different dense area has a good diversity,which can make up for the shortage of the non-maximum suppression algorithm,and then the final result is obtained by combining the output of these two.This paper verifies the effectiveness of the algorithm on two pedestrian datasets with a large number of overlaps by using two different kinds of object detectors respectively.The evaluation results are presented by pricision,recall and F1 value.Experimental results show that the performance of the multi-stage post-processing algorithm proposed in this paper is superior to other post-processing algorithms on different data sets and different target detectors,which proves the effectiveness and generalization of the proposed method.