Rethinking Similar Object Interference in Single Object Tracking

Yipei Wang, Shiyu Hu, Xin Zhao

Abstract: Similar object interference (SOI) problem challenges the single object tracking (SOT) task, leading to the failure of feature-based trackers and subsequent performance degradation. Unfortunately, current generic SOT benchmarks do not effectively tackle this critical challenge, while popular SOT algorithms consistently underestimate the influence they have on tracking performance. To bridge this gap and further enhance the investigation of similar object interference in SOT, we adopt the following viewpoints: (1) By examining the operational principles of mainstream trackers and their performance on representative SOT datasets, we redefine similar objects, taking into account the cognitive bias that exists between trackers and humans when dealing with this challenge. (2) Subsequently, we develop a mining methodology that enables the extraction of the SOI sub-dataset from SOT datasets without relying on human intervention. This methodology comprises two main components: determining the SOI challenge and screening the SOI sequences. The SOI dataset is acquired from representative SOT dataset using our proposed approach, known as SOI2023. This dataset serves as an ideal environment to facilitate the investigation of challenges related to similar object interference. (3) Additionally, we conduct extensive tracking experiments with 20 typical trackers and their variants on SOI2023 and analyze their performance for similar object interference scenes in several dimensions. The experimental results demonstrate the effectiveness of our proposed mining method, while revealing the strengths and weaknesses of current trackers when faced with the challenge of similar object interference. We hope this work can provide inspiration to the tracking community and also provide support and insights for robust tracking under the SOI challenge.