A semi-synthetic dataset of infrared dim small moving targets for detection and segmentation

Infrared dim small moving target detection is a research hotspot in the field of target perception, and it is widely used in security monitoring, visual guidance, reconnaissance and early warning. Recently, infrared dim small target detection algorithms based on deep learning are developing rapidly. These algorithms require a large amount of data for model training and testing. However, it is a complicated and enormous work to obtain infrared dim small moving target data and annotate them precisely in large quantities. Therefore, existing datasets either synthesize targets using Gaussian distribution or sample frames to reduce the workload, which make targets unreal and motion trajectories damaged respectively. To solve the above problems, we proposed a generation adversarial network with the spatio-temporal feature fusion, which can learn the appearance change characteristics of real targets to generate highly simulated infrared targets. Then, combining the rich scenes and various target motion trajectories of existing datasets, we finally provided a large-scale semi-synthetic dataset for infrared dim small moving target detection and segmentation under clutter background. Our dataset covers four scenes (building, vegetation, cloud and water), three cases (one, two and three targets), and provides three annotations (center point, target box and mask), as well as two kinds of evaluation metrics (instance-level and pixel-level), which is conducive to promoting the development of target detection models based on instance and segmentation. Our dataset contains 200 image sequences, 60000 images, which are equally divided into a training set and a test set. It can be used in infrared dim small moving target detection, segmentation and tracking researches.