Deep Object Tracking With Shrinkage Loss.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Data imbalance is a critical issue in deep learning models, particularly in visual object tracking.
• Existing deep regression trackers suffer from imbalanced pixel differences in loss computation, limiting their performance.
• Deep classification trackers face class imbalance due to insufficient positive samples compared to negative ones.

Purpose of the Study:

• To address data imbalance in deep visual object tracking.
• To propose a novel shrinkage loss function to improve tracker performance.
• To enhance the ability of trackers to distinguish target objects from backgrounds.

Main Methods:

• Introduced a novel shrinkage loss function to penalize easy training data, primarily background samples.
• Applied the shrinkage loss to both deep regression and deep classification trackers.
• Validated the proposed method on six benchmark datasets: OTB-2013, OTB-2015, UAV-123, VOT-2016, VOT-2018, and LaSOT.

Main Results:

• The proposed one-stage deep regression tracker with shrinkage loss outperformed state-of-the-art methods, including discriminative correlation filters (DCFs).
• Shrinkage loss generalized well to deep classification trackers, significantly improving performance.
• Replacing binary cross-entropy loss with shrinkage loss led to substantial gains in three baseline trackers, achieving new state-of-the-art results.

Conclusions:

• The novel shrinkage loss effectively mitigates data imbalance issues in visual object tracking.
• The proposed method enhances the discriminative power of both regression and classification-based trackers.
• Shrinkage loss offers a generalizable solution for improving deep visual object tracking performance across various datasets and architectures.