Anhui University | small target tracking: large-scale data sets and baselines

Paper title ： Tiny Object Tracking: A Large-scale Dataset and A Baseline 

Thesis link ：https://arxiv.org/abs/2202.05659

Author's unit ： Anhui University & Pengcheng Laboratory

In practical application, the appearance and characteristics of small objects are weak , And it has attracted more and more attention in common visual tasks such as object detection and segmentation . In order to promote the research and development of micro target tracking , We created a large-scale video data set , It includes 434 A sequence of , Total over 217K frame . Each frame is carefully annotated with high-quality bounding boxes . In data creation , We have considered 12 Challenge attributes to cover a wide range of viewpoints and scene complexity , These attributes are annotated to facilitate attribute based performance analysis . In order to provide a powerful baseline in tiny object tracking , We propose a novel multistage knowledge distillation network （MKDNet）, It carries out three-level knowledge distillation in a unified framework , To effectively enhance the feature representation of tracking small objects 、 Ability to identify and locate . Extensive experiments have been carried out on the proposed data set , It turns out that MKDNet Superiority and effectiveness compared with the most advanced methods .