A Multi-object Detection Sampling Algorithm For Large Scenes

Abstract

Multi-object detection in large scenes aims to find objects in images, which usually contain more than one billion pixels. Based on the concept of dividing and conquering, the state-of-the-art (SOTA) methods slice the super-resolution image into patches first and then lower the image solution to detect objects later. The advantage of this method is that it can adapt quickly to regular detection algorithms. However, a set of parameters needs to be set manually, such as the size of sliding windows and overlap, which is quite hard to fit all scenarios. It may result in a loss of samples located at the boundary of the sliding window and the oversampling of inefficient samples that appear within the overlap. In this paper, we propose a object-oriented image sampling algorithm based on anchor boxes during training and multi-scale pyramids during inference. Inspired by the mature object detection baseline Scale-YOLOv4, we present more tricks to fit large scenes. The accuracy can reach 66%, which is 24 points higher than the CascadeRCNN model of the official backbone network ResNet50. Finally, we have won first place in the PANDA object detection tracking using this method.