An Approach to Streaming Video Segmentation With Sub-Optimal Low-Rank Decomposition

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Video segmentation is crucial for analyzing visual data.
• Existing methods struggle with noise and data corruptions.
• Robust and efficient segmentation remains a challenge.

Purpose of the Study:

• To develop a robust and efficient video segmentation algorithm.
• To suppress the effects of data noises and corruptions.
• To enable streaming video segmentation in unsupervised and interactive scenarios.

Main Methods:

• Proposed a general algorithm: sub-optimal low-rank decomposition (SOLD).
• Employed SOLD to achieve low-rank representation for video segmentation.
• Incorporated a discriminative replication prior into SOLD.
• Developed an efficient inference algorithm based on SOLD for streaming segmentation.
• Utilized constrained normalized-cut with low-rank representation and temporal constraints.

Main Results:

• The SOLD algorithm effectively pursues low-rank representation for video segmentation.
• The discriminative replication prior leverages recurring video patterns.
• The streaming inference algorithm performs unsupervised and interactive segmentation.
• Experiments on VSB100 and SegTrack datasets demonstrate superior performance.
• The approach shows improvements in both accuracy and efficiency.

Conclusions:

• The proposed SOLD-based approach offers robust and efficient video segmentation.
• The method successfully suppresses noise and corruptions.
• It outperforms existing video segmentation techniques in accuracy and efficiency.
• Enables effective streaming video segmentation.