Self-Supervised Video Hashing With Hierarchical Binary Auto-Encoder.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Existing video hashing methods often isolate stages like frame pooling, relaxed learning, and binarization.
• This isolation leads to inadequate exploration of temporal frame order in joint binary optimization, causing significant information loss.
• There is a need for methods that effectively capture temporal video information for accurate retrieval.

Purpose of the Study:

• To propose a novel unsupervised video hashing framework, Self-Supervised Video Hashing (SSVH).
• To address the challenges of designing an encoder-decoder architecture for video binary code generation.
• To enhance the accuracy of video retrieval using generated binary codes.

Main Methods:

• Developed a hierarchical binary auto-encoder to model temporal dependencies at multiple granularities.
• Implemented an end-to-end learning-to-hash approach to embed videos into binary codes efficiently.
• Encouraged binary codes to reconstruct both visual content and neighborhood structure of videos.

Main Results:

• The proposed SSVH method significantly outperforms state-of-the-art unsupervised video retrieval techniques.
• Achieved current best performance on unsupervised video retrieval tasks across two real-world datasets.
• Demonstrated effective capture of temporal video nature through joint binary optimization.

Conclusions:

• SSVH offers a superior approach to unsupervised video hashing by integrating temporal information.
• The framework effectively generates binary codes for accurate and efficient video retrieval.
• SSVH represents a significant advancement in learning-to-hash for video analysis.