SiamDF: Tracking training data-free siamese tracker.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Siamese tracking algorithms have advanced significantly, largely due to extensive training datasets.
• The specific role of large datasets in enhancing siamese tracker effectiveness remains under-explored.

Purpose of the Study:

• To analyze the impact of training data on siamese tracker learning from an optimization viewpoint.
• To develop a novel data-free siamese tracking algorithm that leverages insights from data analysis.

Main Methods:

• Investigated the function of training data in background suppression and target representation refinement.
• Introduced SiamDF, a data-free algorithm requiring only a pre-trained backbone.
• Enhanced background suppression by isolating target regions and employing inverse transformations for aspect ratio consistency.
• Improved center displacement prediction by mitigating spatial stride deviations.

Main Results:

• SiamDF demonstrates effective background suppression and refined target representation without additional training.
• The algorithm maintains target state aspect ratio and improves center displacement prediction.
• Experimental results show SiamDF achieves competitive performance against supervised and unsupervised methods on popular benchmarks.

Conclusions:

• Large training datasets are crucial for background suppression in siamese tracking.
• The proposed SiamDF algorithm offers a viable data-free approach to siamese tracking, eliminating the need for fine-tuning or online updates.
• SiamDF achieves state-of-the-art performance, highlighting the potential of data-efficient learning strategies.