Rate-Performance-Loss Optimization for Inter-Frame Deep Feature Coding From Videos.

Area of Science:

• Computer Vision
• Machine Learning
• Data Compression

Background:

• The proliferation of video data from surveillance and mobile devices necessitates efficient transmission for cloud-based big data analysis.
• Deep convolutional neural networks (CNNs) extract high-performance features, but compressing these deep video features without performance loss remains a challenge.

Purpose of the Study:

• To propose a novel framework for high-efficiency deep feature coding (DFC) of videos.
• To address the open problem of compressing deep video features while preserving analysis and retrieval performance.

Main Methods:

• Introduced a deep feature coding (DFC) framework with three feature types (I-feature, P-feature, S-feature) within groups-of-features (GOFs).
• Developed sequential and adaptive prediction structures for features within GOFs.
• Proposed a rate-performance-loss optimization model for P-feature residual coding.
• Created the VFC-1M dataset with 1 million visual objects from real-world surveillance videos for evaluation.

Main Results:

• The DFC framework significantly reduces the bitrate of deep video features.
• Video retrieval accuracy is maintained despite substantial bitrate reduction.
• The VFC-1M dataset facilitates evaluation of video feature coding methods.

Conclusions:

• The proposed DFC framework offers an effective solution for compressing deep video features.
• This approach enables efficient transmission of video data for large-scale cloud analysis and retrieval.
• The method balances bitrate reduction with the preservation of critical analysis and retrieval performance.