A Joint Compression Scheme of Video Feature Descriptors and Visual Content

Area of Science:

• Computer Vision
• Multimedia Signal Processing
• Information Theory

Background:

• Mobile visual retrieval demands efficient feature descriptor compression.
• Transmitting only descriptors limits applications due to missing visual content.
• A hybrid approach is needed for joint content and feature compression.

Purpose of the Study:

• Investigate a content-plus-feature coding scheme for next-generation video compression.
• Enable high-efficiency coding of both features and visual content by exploiting their interactions.
• Optimize visual retrieval performance and video coding efficiency simultaneously.

Main Methods:

• Utilize video stream structure and motion information for compact feature descriptor representation.
• Propose a novel rate-accuracy optimization technique for feature coding.
• Employ compressed feature data for improved video coding via feature matching-based affine motion compensation.

Main Results:

• Achieve significant bitrate reduction for both feature descriptors and video frames.
• Maintain state-of-the-art visual retrieval performance.
• Demonstrate the effectiveness of the joint compression framework through extensive simulations.

Conclusions:

• The proposed joint compression framework effectively balances bitrate reduction and retrieval accuracy.
• This approach facilitates the widespread adoption of mobile visual retrieval systems.
• It paves the way for future video compression systems optimized for visual retrieval.