Motion Segmentation & Multiple Object Tracking by Correlation Co-Clustering.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Computer vision models typically focus on either low-level pixel grouping or high-level object detection and tracking.
• Integrating these bottom-up and top-down approaches presents a significant challenge.

Purpose of the Study:

• To propose a unified framework for simultaneously addressing low-level motion segmentation and high-level object tracking.
• To demonstrate the efficacy of a co-clustering approach for this joint problem.

Main Methods:

• Formulating the joint problem as a co-clustering task, leveraging existing algorithms.
• Combining bottom-up motion segmentation (grouping point trajectories) with top-down multiple object tracking (clustering bounding boxes).

Main Results:

• The proposed co-clustering method significantly improves motion segmentation on the FBMS59 benchmark.
• The approach achieves state-of-the-art results on multiple object tracking benchmarks (MOT15, MOT16, MOT17).

Conclusions:

• Solving the joint problem of low-level grouping and high-level tracking via co-clustering is beneficial for both tasks.
• This unified approach offers a principled and effective solution for integrated computer vision tasks.