TRECVID 2009 of MCG-ICT-CAS

Juan Cao, Yong-Dong Zhang, Bai-Lan Feng, Lei Bao, Ling Pang, Jin-Tao Li, Ke Gao, Xiao Wu, Hon-Ttao Xie, Wei Zhang, Zhen-Dong Mao

Abstract

This paper details the participation of the MCG-ICT-CAS team in the TRECVID 2009 challenge, focusing on two primary domains: interactive video search and content-based copy detection (CBCD). We introduce VideoMap, a visualization-driven search interface that utilizes multi-strategy feedback and community mining to bridge the semantic gap. Additionally, we present a high-performance CBCD system optimized with GPU acceleration and spatial-temporal consistency checks to identify distorted video copies efficiently. Experimental results show that our multi-strategy search approach achieved a Mean Average Precision (MAP) of 0.186, while our GPU-based interest point detector provided a 10-20x speedup over traditional CPU methods.

Introduction

As digital video libraries grow exponentially, the "semantic gap"—the disconnect between low-level visual features and high-level human concepts—remains a significant hurdle for effective retrieval. Simultaneously, the industry faces challenges in Content-Based Copy Detection (CBCD), as pirated content is often modified (e.g., cropped, flipped, or embedded as a "Picture-in-Picture") to evade automated detection. This work details the MCG-ICT-CAS team's efforts to enhance search intuitiveness and copy detection robustness for the TRECVID 2009 challenge.

Interactive Search: The VideoMap System

To improve the efficiency of human-in-the-loop video search, we developed VideoMap. Unlike traditional list-based search results, VideoMap organizes the video database into a 2D spatial layout where similar shots are clustered together.

System Workflow and Multi-Strategy Feedback

The VideoMap system integrates several advanced retrieval strategies to assist the user:

  1. Visual-based Feedback: Leverages keyframe similarity to find visually related content.
  2. Concept-based Feedback (GMSSCD): Employs "Graph-Based Multi-Space Semantic Correlation Diffusion" to expand initial queries into a broader set of semantically related concepts.
  3. Community-based Feedback: Models video shots as a network, identifying "communities" that share latent semantic topics.
The VideoMap system workflow, integrating multi-modal similarity, community mining, and user feedback for re-ranking.

A key feature of VideoMap is Active Annotation, which allows users to label entire regions of the map as relevant or irrelevant simultaneously. The system can intelligently auto-switch between these strategies based on user behavior and the quality of the retrieved results.

Content-Based Copy Detection (CBCD)

Our CBCD system is designed to identify video copies even under severe distortions. The system consists of four specialized modules (Global, Local, PIP, and Flip) and leverages several key technologies:

  • Block Gradient Histogram (BGH): A robust frame descriptor invariant to blur, noise, and color shifts.
  • GPU-Accelerated Harris-Hessian (H-H): A novel interest point detector implemented on the GPU, significantly reducing the computational overhead of feature extraction.
  • Time Sequence Consistency (TSC): To ensure temporal alignment, we treat video matching as a sequence problem, utilizing a "maximum flow" graph algorithm to find the most consistent temporal path between the query and the database.

Experimental Evaluation and Results

The systems were evaluated using the TRECVID 2009 benchmarks for interactive search and copy detection.

Search Performance

The multi-strategy feedback approach achieved a peak MAP of 0.186. Notably, "Run 9," which fused visual and temporal similarity, significantly outperformed visual-only baselines, proving that video retrieval must treat video as a continuous medium rather than a collection of still photos.

Average precision across TRECVID 2009 topics, comparing the MCG-ICT-CAS run score (dots) against median and best results.

Copy Detection Efficiency

The GPU-based Harris-Hessian detector achieved a 10-20x speedup compared to CPU implementations, processing a 640x480 image in approximately 6.3ms. The system demonstrated high robustness, successfully detecting copies even when they occupied only 20% of the frame (PIP) or were vertically flipped.

Conclusion

The MCG-ICT-CAS contributions to TRECVID 2009 demonstrate that multi-strategy feedback and visualization-driven interfaces like VideoMap can significantly enhance interactive search performance. Furthermore, GPU acceleration and temporal consistency checks are vital for making robust copy detection feasible at scale. Future work will focus on refining concept detectors to further narrow the semantic gap and optimizing GPU memory management for even greater efficiency.

Summary updated

BibTeX
@inproceedings{cao2009trecvid,
  author = {Juan Cao and Yong-Dong Zhang and Bai-Lan Feng and Lei Bao and Ling Pang and Jin-Tao Li and Ke Gao and Xiao Wu and Hon-Ttao Xie and Wei Zhang and Zhen-Dong Mao},
  title = {{TRECVID 2009 of MCG-ICT-CAS}},
  booktitle = {TRECVID},
  year = {2009},
}
Figure