Abstract
Visual object tracking is frequently challenged by "model drift," where a tracker gradually loses the target due to interference from background noise or similar-looking objects. This paper introduces SPSTracker, a framework designed to mitigate this issue by explicitly regularizing the tracking response map. By incorporating two lightweight modules—Boundary Response Truncation (BRT) and Peak Response Pooling (PRP)—SPSTracker suppresses secondary "sub-peaks" and aggregates scattered signals into a single, sharp peak. Evaluated on multiple benchmarks, SPSTracker achieves state-of-the-art performance while maintaining real-time speeds of 35 FPS.
Introduction
In modern discriminative tracking, the location of a target is typically determined by the maximum value in a generated response map. Ideally, this map should exhibit a single, clear Gaussian-like peak centered on the target. However, real-world scenarios involving occlusions, illumination changes, or distractor objects often lead to the emergence of "sub-peaks"—false local maxima that can mislead the tracker. If these sub-peaks are used for online model updates, the tracker's internal representation becomes corrupted, leading to progressive model drift and eventual failure.
Methodology
SPSTracker is built upon the ATOM tracker architecture and introduces a "rectified online learning" strategy focused on output regularization.
Boundary Response Truncation (BRT)
The BRT module is a simple yet effective clipping operation. It identifies response values located far from the estimated peak and sets them to zero. This truncation reduces the variance of the feature response and prevents background noise from being incorporated into the model update process.
Peak Response Pooling (PRP)
PRP is designed to handle scattered signals within the target region. Inspired by pooling operations in object detection, PRP performs horizontal and vertical maximization across the response map. This process "squashes" multiple sub-peaks and aligns their energy toward the geometric center of the target, ensuring the strongest signal is concentrated at the correct location.
Feature Fusion and Architecture
The framework utilizes a ResNet-18 backbone, specifically fusing features from Block3 and Block4. This combination allows the tracker to leverage both high-level semantic information and low-level spatial details (like color and edges) for precise localization.
Experimental Results
The authors evaluated SPSTracker on six major benchmarks, including OTB2013/2015, VOT2016/2018, and the high-frame-rate NFS dataset.
- VOT2018: SPSTracker achieved an Expected Average Overlap (EAO) of 0.434, outperforming the baseline ATOM (0.401) and SiamMask (0.380).
- OTB2015: The tracker reached a success score of 0.692, representing a 2.1% improvement over the baseline.
- Efficiency: Despite the added modules, SPSTracker maintains a real-time processing speed of 35 FPS on a standard GTX 1080ti GPU.
Conclusion
SPSTracker demonstrates that regularizing the distribution of the tracking response is a highly efficient way to improve robustness against model drift. By suppressing sub-peaks through the lightweight BRT and PRP modules, the tracker achieves superior accuracy without sacrificing real-time performance. Future work may explore adaptive truncation boundaries to further optimize the balance between noise suppression and foreground preservation.
Summary updated