Abstract
<p indent=0mm>The performance of object detection algorithms depends on both dataset distribution and network design of feature extraction. Starting from these two points, we firstly explore the potential inherent reasons that lead to low detection accuracy of small object by analyzing the distribution of object attributes at various scales in the COCO 2017 dataset, and propose copy and paste (CP) module accordingly, which adjusts the distribution of small object offline, on the one hand, upsampling the pictures containing small objects, on the other hand, copying and pasting the small objects in the pictures. Then, to further improve network feature extraction ability, inspired by the idea of curriculum learning (CL), we propose CL layer, which uses ground truth labels to guide the learning process, and CL factor to control the learning intensity, the features of objects are enhanced to facilitate network feature extraction. We deploy the CP module on the COCO 2017 dataset and embed the CL layer in the CenterNet network to conduct multiple sets of comparative experiments, and use average detection accuracy, small object detection accuracy, medium object detection accuracy, and large object detection accuracy as evaluation indicators. The experimental results prove the effectiveness of CP module and CL layer.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have