Improving Object Tracking by Added Noise and Channel Attention

Mustansar Fiaz, Arif Mahmood, Ki Yeol Baek, Sehar Shahzad Farooq, Soon Ki Jung2020 · DOI

TL;DR

This paper introduces the IRCA-Siam tracker, which improves visual object tracking by adding noise to training data for better generalization and incorporating a channel attention mechanism to focus on more informative features. The proposed method achieves state-of-the-art performance across several challenging benchmarks.

What Problem Does This Paper Address?

Visual object tracking (VOT) is a crucial task in computer vision, aiming to follow a target object through a video sequence. While deep learning, particularly Siamese networks, has significantly advanced VOT, these models often struggle with generalization and can overfit to training data, especially when dealing with diverse real-world challenges like scale variations, occlusion, and illumination changes. Existing Siamese trackers, despite their efficiency, can also suffer from low-fidelity target representation. This paper addresses the challenge of improving the generalization ability and robustness of Siamese-based trackers to achieve better performance in complex scenarios.

Methods

The core of the proposed IRCA-Siam tracker lies in two main innovations:

  • Input Noise Regularization: The authors introduce additive Gaussian noise to the input images (both template and search regions) during the training phase. This acts as a form of data augmentation and regularization, forcing the network to learn more robust features that are less sensitive to noise and variations, thereby improving generalization.
  • Channel Attention Mechanism: To further enhance feature discrimination, a channel attention module is integrated into the template branch. This module learns to assign different weights to feature channels, emphasizing those that are more informative for target localization and discrimination while suppressing less relevant ones. This is achieved by fusing global average pooling and global maximum pooling features.
  • Early Feature Fusion: Low-level features from both noisy and clean input channels are fused early in the network. This fusion aims to create a more robust target representation that benefits from both the enhanced robustness of noisy features and the clarity of clean features, leading to improved target localization.

The framework builds upon the SiameseFC architecture, modifying its training process and incorporating the attention mechanism.

Datasets

The authors trained their model on the GOT-10K dataset, which contains over 10,000 video sequences.

For evaluation, they used several widely recognized benchmark datasets:

  • OTB2013
  • OTB2015
  • TC128
  • UAV123
  • VOT2016
  • VOT2017

Key Insights and Ideas

  • Noise as a Regularizer: The paper demonstrates that strategically adding noise during training, rather than relying solely on dropout, can be a highly effective way to improve a deep tracker's generalization capabilities and robustness to real-world noise.
  • Channel Attention for Focus: The channel attention mechanism proves crucial in allowing the tracker to dynamically focus on the most discriminative parts of the target's appearance, leading to better distinction from background clutter.
  • Early Fusion for Robustness: Fusing features from both noisy and clean inputs early in the network is shown to be beneficial for creating a more resilient target representation, especially in challenging tracking scenarios.
  • Synergy of Techniques: The combination of input noise regularization and channel attention yields a synergistic effect, leading to performance gains that surpass either technique applied in isolation.

Key Results

The IRCA-Siam tracker demonstrated superior performance across multiple benchmarks, outperforming 30 other state-of-the-art trackers.

  • OTB2015: Achieved 62.5% precision and 83.5% success, showing significant gains over the baseline SiamFC.
  • VOT2016: Achieved the best accuracy (0.56) among compared methods.
  • VOT2017: Achieved the highest overlap score (0.52) and best accuracy.
  • Real-time Performance: The tracker operates at 77 FPS, making it a real-time capable solution.
  • Challenge Performance: IRCA-Siam showed strong performance across various challenges, including fast motion, motion blur, deformation, and occlusion, often ranking first or second.

Limitations and Future Work

  • Low-Resolution and Background Clutter: The paper notes that IRCA-Siam did not perform as well on videos with low resolution or significant background clutter, though it still ranked second in these cases.
  • Further Exploration of Noise Layers: While the paper investigates adding noise layers before convolutional layers, it notes that this approach sometimes degraded performance compared to input noise. Further research could explore optimal placement and types of noise layers.
  • Generalization to Unseen Scenarios: While the proposed methods improve generalization, the inherent challenge of tracking objects in entirely novel and unseen scenarios remains an area for future exploration.

The authors suggest that their approach of combining input noise regularization and channel attention provides a strong foundation for future work in developing more robust and accurate visual object trackers.

Summary updated

BibTeX
@misc{fiaz2020improving,
  author = {Mustansar Fiaz and Arif Mahmood and Ki Yeol Baek and Sehar Shahzad Farooq and Soon Ki Jung},
  title = {{Improving Object Tracking by Added Noise and Channel Attention}},
  year = {2020},
  doi = {10.3390/s20133780},
}
Figure