Event Stream-based Visual Object Tracking: HDETrack V2 and A High-Definition Benchmark

TL;DR

This work tackles visual object tracking with event cameras by introducing HDETrack V2, a teacher–student Transformer framework that transfers knowledge from a multi-modal RGB+Event teacher to a pure event-based student via hierarchical knowledge distillation. The approach enhances temporal modeling with a temporal Fourier transform and enables robust target adaptation through test-time tuning and adaptive search regions. A new large-scale, high-definition EventVOT dataset (1280×720) is proposed to benchmark event-based tracking across diverse categories and challenging conditions. Empirical results show that HDETrack V2 outperforms strong baselines on EventVOT and other event datasets, validating the effectiveness of cross-modal knowledge transfer, online adaptation, and high-resolution event data for real-time tracking applications.

Abstract

We then introduce a novel hierarchical knowledge distillation strategy that incorporates the similarity matrix, feature representation, and response map-based distillation to guide the learning of the student Transformer network. We also enhance the model's ability to capture temporal dependencies by applying the temporal Fourier transform to establish temporal relationships between video frames. We adapt the network model to specific target objects during testing via a newly proposed test-time tuning strategy to achieve high performance and flexibility in target tracking. Recognizing the limitations of existing event-based tracking datasets, which are predominantly low-resolution, we propose EventVOT, the first large-scale high-resolution event-based tracking dataset. It comprises 1141 videos spanning diverse categories such as pedestrians, vehicles, UAVs, ping pong, etc. Extensive experiments on both low-resolution (FE240hz, VisEvent, FELT), and our newly proposed high-resolution EventVOT dataset fully validated the effectiveness of our proposed method. Both the benchmark dataset and source code have been released on https://github.com/Event-AHU/EventVOT_Benchmark

Figures (12)

• Figure 1: (a). Comparison between our newly proposed EventVOT and other event-based tracking datasets; (b). RGB-Event based multi-modal tracking framework; (c). Pure event-based tracking framework; (d). Our newly proposed Test-Time Tuning based event-tracking framework.
• Figure 2: An overview of our proposed Hierarchical Knowledge Distillation Framework for Event Stream based Tracking, termed HDETrack V2. It contains the teacher and student Transformer networks which take multi-modal/multi-view and event data only as the input respectively. Both networks share an identical architecture, i.e., tracking using a unified Transformer backbone network similar to CEUTrack tang2022coesot and OSTrack ye2022Ostrack. Specifically, we extract the template and search patches from both RGB and event inputs, generating feature embeddings through a projection layer. These embeddings are then passed through a stack of Transformer layers that form the teacher network. The output of this network feeds into the tracking head for target object localization. Meanwhile, the student network is designed for efficient tracking using event stream only. It is trained with tracking loss functions and benefits from knowledge distillation from the teacher Transformer network. Our tracker achieves a better tradeoff between accuracy and model complexity.
• Figure 3: The Test Time Tuning (TTT) strategy employed during the inference phase. The template frames are augmented based on the sparsity of the initial event streams and then fused within the search region to yield a variety of response maps. We further enhance the tracker's efficacy through the integration of LoRA specifically tailored for the testing phase. It is worth noting that, to maintain alignment between the TTT stage and the foundational training stage, we achieve it by using the tracking results of the initial few frames in a video as the pseudo labels for self-supervised learning. For more detailed implementation, please refer to section \ref{['subsec:TTTT']}.
• Figure 4: Representative samples of our proposed EventVOT dataset. The $1^{th}$ column is the 3D event point stream and the $2^{th}$-$4^{th}$ columns are sampled event images. $5^{th}$-$7^{th}$ columns are more samples of our EventVOT dataset.
• Figure 5: Distribution visualization of challenging factors, category of the target object, and bounding box.