The Power of Tiling for Small Object Detection

Abstract

Deep neural network based techniques are state-of-the-art for object detection and classification with the help ofthe development in computational power and memory ef-ficiency. Although these networks are adapted for mobileplatforms with sacrifice in accuracy; the resolution increasein visual sources makes the problem even harder by raisingthe expectations to leverage all the details in images. Real-time small object detection in low power mobile devices hasbeen one of the fundamental problems of surveillance ap-plications. In this study, we address the detection of pedes-trians and vehicles onboard a micro aerial vehicle (MAV)with high-resolution imagery. For this purpose, we exploitPeleeNet, to our best knowledge the most efficient networkmodel on mobile GPUs, as the backbone of an SSD networkas well as 38x38 feature map in the earlier layer. After illus-trating the low accuracy of state-of-the-art object detectorsunder the MAV scenario, we introduce a tiling based ap-proach that is applied in both training and inference phases.The proposed technique limits the detail loss in object de-tection while feeding the network with a fixed size input.The improvements provided by the proposed approach areshown by in-depth experiments performed along Nvidia Jet-son TX1 and TX2 using the VisDrone2018 dataset.

Tasks

Reproductions