Single object tracking based on OSTrack and NAS head
Conghao Li, Haofeng Wang, Pengjie Liu · 2022
Neural architecture search (NAS) has attracted much attention in recent years. Especially, Differentiable Architecture Search (DARTS) has become a mainstream NAS algorithm because of its differentiable architecture hyper-parameter and low time complexity, but few people apply it in the field of object tracking due to its high memory complexity. This paper redesigns the cell structure according to the characteristics of the target tracking head network and defines a simplified search space. We propose a NASHT model to replace the three head networks of OSTrack with the modified DARTS cell, which automatically searches for a target structure suitable for its task. We use two benchmarks, PTB-TIR and LSOTB-TIR, to train and evaluate the model. The accuracy of the proposed model reached 86.5% and the success rate reached 69.0% on the PTB-TIR benchmark. Compared with OSTrack, the accuracy has increased by 1.5%, and the success rate has a 0.8% improvement. The accuracy of the proposed model on the LSOTB-TIR reached 84.9% and the success rate was 71.4%, which achieved comparable effects to OSTrack while performing better in situations such as motion blur, low resolution, and aspect ratio changes.