NCSiam: Reliable Matching via Neighborhood Consensus for Siamese-Based Object Tracking

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Accurate visual object tracking is crucial for many AI applications.
• Current Siamese-based trackers struggle with erroneous matches due to limitations in similarity map generation.
• Existing methods often discard valuable appearance and semantic information during tracking.

Purpose of the Study:

• To propose a novel Siamese-based tracker, NCSiam, that refines similarity maps using neighborhood consensus constraints.
• To enhance the discriminability between tracking targets and distractors by analyzing larger scene contexts.
• To leverage semantic information and improve representation learning for more robust object tracking.

Main Methods:

• Developed a neighborhood consensus constraint-based Siamese tracker (NCSiam).
• Introduced a 4D convolution-based multi-level similarity refinement (MLSR) strategy to analyze neighborhood consensus patterns.
• Incorporated an appearance affinity decoder (AAD) to utilize semantic information and a task-specific disentanglement (TSD) module for decoupled embeddings.

Main Results:

• NCSiam demonstrates significant improvements in visual object tracking accuracy.
• The MLSR strategy effectively refines similarity maps by analyzing neighborhood consensus.
• AAD and TSD modules enhance the tracker's ability to distinguish targets from distractors and utilize semantic information.

Conclusions:

• NCSiam offers a novel and effective approach to visual object tracking by addressing limitations in similarity map refinement.
• The proposed neighborhood consensus constraint significantly enhances tracking robustness and accuracy.
• Experimental results across six challenging benchmarks validate the superiority of the NCSiam method.