Exploring Effective Methods to Improve the Performance of Tiny Object Detection

In this paper, we present our solution of the 1st Tiny Object Detection (TOD) Challenge. The purpose of the challenge is to detect tiny person objects (2–20 pixels) in large-scale images. Due to the extreme small object size and low signal-to-noise ratio, the detection of tiny objects is much more challenging than objects in other datasets such as COCO and CityPersons. Based on Faster R-CNN, we explore some effective and general methods to improve the detection performance of tiny objects. Since the model architectures will not be changed, these methods are easy to implement. Accordingly, we obtain the 2nd place with the \(A P_{50}^{{\text {tiny}}}\) score of 71.53 in the challenge.