Abstract
Visual Object Tracking (VOT) faces significant hurdles in low-resolution (LR) environments where target details are sparse. This work introduces a Siamese network framework enhanced by Random Noise Feature Fusion and a Salient Feature Attention mechanism. The former acts as a robust regularizer by training on noise-infused features, while the latter utilizes a Convolutional Block Attention Module (CBAM) to focus on critical semantic information. Evaluated on the OTB100 benchmark, the tracker achieves state-of-the-art performance with a success score of 0.685 and a precision of 0.937, while operating at a real-time speed of 68 FPS.
Introduction
Visual Object Tracking is a fundamental task in computer vision, yet its performance degrades sharply in low-resolution (LR) scenarios. Such conditions are prevalent in practical applications like long-range security monitoring, low-cost UAV surveillance, and remote sensing. In these contexts, targets often lack the fine-grained spatial details—such as sharp edges and textures—that traditional trackers rely on. Standard Siamese trackers often overfit to limited data or fail to distinguish targets from noisy backgrounds. This paper addresses these challenges by proposing a network that is inherently robust to signal degradation and sparse target information.
Proposed Methodology
The proposed framework enhances the Siamese Fully Convolutional (SiamFC) architecture through two key innovations: Random Noise Feature Fusion and a Salient Feature Attention Network.
Random Noise Feature Fusion
To improve generalization and prevent overfitting to clean imagery, the authors introduce a noise-infused branch during training. As shown in Figure 1, the network processes both original and noised versions of the target ($z$) and search ($x$) images. Various noise types, including Gaussian, salt and pepper, Poisson, and speckle, are utilized. These noised features are fused with original features via element-wise summation, forcing the network to learn structural representations that are invariant to pixel-level fluctuations.
Salient Feature Attention Network
To prioritize informative regions, a Convolutional Block Attention Module (CBAM) is integrated. This module applies both channel and spatial attention to high-level semantic features extracted from the network's deeper layers. By weighting these "salient" features more heavily, the model can better differentiate the target from complex backgrounds, even when visual details are blurry. The final response map is generated through cross-correlation of these enhanced feature maps.
Experimental Results
The model was trained on standard target and search image pairs and evaluated on the OTB100 benchmark, focusing on sequences with low-resolution attributes.
Quantitative Performance
The proposed tracker achieved a Success score of 0.685 and a Precision score of 0.937. These results surpass several state-of-the-art (SOTA) models, including TransT and AiATrack. In terms of efficiency, the tracker maintains a high processing speed of 68 FPS. While slightly slower than the baseline SiamFC (86 FPS), it significantly outperforms Transformer-based models like TransT (50 FPS) and AiATrack (38 FPS) in real-time capability.
Qualitative Analysis
Qualitative evaluations demonstrate the robustness of the proposed method. In challenging sequences such as "walking2" or tracking a panda, the model consistently maintains a lock on the target where competing trackers often drift or fail due to background clutter or low visibility.
Conclusion
This paper demonstrates that enhancing a Siamese network with noise-based regularization and salient attention significantly improves tracking performance in low-resolution scenarios. By prioritizing semantic information over potentially unreliable spatial details, the proposed tracker achieves a superior balance of accuracy and real-time speed. Future work will explore the optimal placement of attention mechanisms and the tracker's performance in extremely low-resolution conditions.
Summary updated