Learning Dynamic Compact Memory Embedding for Deformable Visual Object Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Template-based trackers excel in efficiency and accuracy but struggle with target deformation and state estimation errors.
• Segmentation-based trackers improve deformable object tracking via per-pixel matching but lack discrimination against challenging factors like distractors and appearance changes.

Purpose of the Study:

• To enhance the discrimination capabilities of segmentation-based visual tracking methods.
• To improve the tracking of deformable objects by addressing limitations in existing approaches.

Main Methods:

• A dynamic compact memory embedding technique is proposed, initialized with target features from the first frame.
• Online updates to the memory embedding incorporate current target features correlated with existing memory.
• A weighted point-to-global matching strategy is employed for pixelwise feature correlation to capture detailed deformation information.

Main Results:

• The proposed method demonstrates superiority over recent trackers on six challenging benchmarks (VOT2016, VOT2018, VOT2019, GOT-10K, TrackingNet, LaSOT).
• The tracker outperforms state-of-the-art segmentation-based trackers (D3S, SiamMask) on the DAVIS2017 benchmark.
• The dynamic memory embedding effectively enhances target discrimination against background clutter and similar distractors.

Conclusions:

• The dynamic compact memory embedding technique significantly improves segmentation-based visual tracking, particularly for deformable objects.
• The method offers enhanced robustness against challenging factors, leading to superior tracking accuracy and discrimination.
• The approach represents a notable advancement in visual tracking, providing a more reliable solution for complex scenarios.