Tracking Small and Fast Moving Objects: A Benchmark

Zhewen Zhang, Fuliang Wu, Yuming Qiu, Jingdong Liang, Shuiwang Li

Abstract

While deep learning has significantly advanced generic object tracking, performance often degrades when targets are small and move at high velocities. This paper introduces the Tracking Small and Fast Moving Objects (TSFMO) benchmark, a specialized dataset comprising 250 sequences across 26 sports categories. We also present S-KeepTrack, a baseline tracker designed to preserve fine-grained spatial details by fusing low-level and high-level features. S-KeepTrack outperforms 20 state-of-the-art trackers on TSFMO, though the overall low performance across all models highlights the ongoing challenges in this domain.

1. Introduction

Current object tracking research primarily focuses on datasets like LaSOT or TrackingNet, where targets are typically large and move with predictable dynamics. However, in real-world applications such as sports broadcasting, objects like tennis balls, shuttlecocks, or pucks present unique difficulties. These objects are characterized by limited visual information, significant motion blur, and abrupt changes in trajectory. To address this, we propose the TSFMO benchmark and a novel tracking architecture, S-KeepTrack, specifically optimized for these high-dynamic scenarios.

2. The TSFMO Benchmark

The TSFMO dataset consists of 250 high-quality video sequences featuring 26 different sports (e.g., badminton, snooker, ice hockey). Each frame is manually annotated with axis-aligned bounding boxes and categorized by 12 distinct attributes, including "Fast Motion," "Motion Blur," and "Low Resolution."

Quantitatively, TSFMO differs drastically from traditional benchmarks. The average target size in TSFMO is only 0.51% of the frame area, compared to 6.73% in the OTB benchmark. Furthermore, the relative speed of objects in TSFMO is nearly 30 times higher than those in OTB. These statistics underscore the specialized nature of the dataset and the difficulty it poses to standard tracking algorithms.

3. S-KeepTrack: Fusing Spatial and Semantic Features

Standard deep trackers often rely on high-level semantic features extracted from deep layers of a convolutional neural network. While these features provide robust object identity, their large receptive fields often "swallow" or overlook tiny objects, leading to localization errors.

S-KeepTrack addresses this by modifying the KeepTrack architecture to incorporate a dual-feature fusion strategy:

  • Parallel Feature Extraction: The model extracts both low-level features from early layers (preserving spatial resolution) and high-level features from deeper layers (providing semantic context).
  • Dual Association: Two parallel candidate association networks process these features independently.
  • Weighted Fusion: The final assignment is determined by a weighted sum of the similarity matrices: $S^* = \omega S + (1 - \omega)\bar{S}$.

This approach ensures that the tracker maintains the fine-grained spatial cues necessary to pinpoint a tiny ball while still benefiting from the discriminative power of deep features.

4. Experimental Results and Discussion

We evaluated 20 state-of-the-art trackers (including TransT, DiMP, and SiamRPN++) on the TSFMO benchmark. S-KeepTrack achieved the highest performance with a Precision (PRC) of 0.437 and an Area Under Curve (AUC) of 0.255.

4.1 The "Deep Feature" Trap

A key finding from the ablation studies is that for tiny objects, deeper features are not necessarily better. When setting $\omega=0.0$ (using only low-level features), the model actually outperformed the configuration using only high-level features ($\omega=1.0$). This suggests that the loss of spatial resolution in deep layers is a primary bottleneck for tracking small objects.

4.2 Challenges in Fast Motion

"Fast Motion" was identified as the most prevalent and difficult attribute. It appeared in nearly every sequence and was the leading cause of tracking failure. Even the best-performing models struggled to maintain lock during rapid direction changes or extreme motion blur, as seen in complex scenarios like "polo_4" or "basketball_14."

5. Conclusion

The TSFMO benchmark reveals a significant performance gap in current tracking technology. While S-KeepTrack provides a stronger baseline by effectively utilizing low-level features, the low absolute AUC scores across all tested models indicate that tracking small, fast-moving objects remains an open challenge. Future research should focus on developing architectures that can better handle extreme dynamics and limited visual cues in high-dynamic environments.

Summary updated

BibTeX
@inproceedings{zhang2022tracking,
  author = {Zhewen Zhang and Fuliang Wu and Yuming Qiu and Jingdong Liang and Shuiwang Li},
  title = {{Tracking Small and Fast Moving Objects: A Benchmark}},
  booktitle = {ACCV},
  year = {2022},
  doi = {10.1007/978-3-031-26293-7\_33},
}
Figure