National Institute of Informatics, Japan at TRECVID 2009

Duy-Dinh Le, Sébastien Poullot, Xiaomeng Wu, Michael Nett, Michael E. Houle, Shinichi Satoh, Michel CrucianuTRECVID 2009 · OpenReview

Introduction

As digital video libraries expand, the need for efficient content categorization, retrieval, and copyright protection becomes critical. Researchers from the National Institute of Informatics (NII), Japan, addressed these challenges in their TRECVID 2009 submission. Their work focuses on three core areas: High-Level Feature Extraction (HLF), video search, and Content-Based Video Copy Detection (CBVCD). A primary goal was resolving the "scalability vs. accuracy" trade-off, particularly in copy detection, where robust local features are often too computationally expensive for large-scale applications.

Methodology

The HLF task involved identifying semantic concepts within video frames using a fusion of global features (color moments, histograms, edge orientation) and local SIFT descriptors. SVM classifiers with RBF kernels were primarily used, though the team also explored faster K-Nearest Neighbor (KNN) classifiers using a SASH indexing structure. The search system integrated these visual detectors with textual metadata matching, employing "Concept Selection" to rank results based on relevant semantic categories found within queries.

Content-Based Video Copy Detection: The "Glocal" Approach

The technical centerpiece of the submission is the "Glocal" signature method. This approach seeks a middle ground between fast global features and robust local features.

Overview of the video copy detection pipeline, integrating glocal descriptors and geometric indexing.

The pipeline utilizes an "Improved Harris Corner Detector" to extract compact 20-dimensional descriptors. To maintain robustness against "Picture-in-Picture" effects and cropping, local features are grouped into "triplets" based on spatial proximity. This preserves the short-range geometric structure of the frame, ensuring that even if a large portion of the frame is modified, the original content remains identifiable.

Spatial partitioning and binary embedding of local features into fixed-length Glocal signatures.

As illustrated in the figure above, these features are embedded into fixed-length binary vectors. This binary representation allows for extremely fast matching using the Dice coefficient, enabling the system to scale effectively to large datasets. To further boost detection confidence, the team incorporated audio fingerprinting using Haitsma's method.

Key Results and Discussion

The team's best HLF fusion run achieved a Mean Average Precision (MAP) of 0.110. A key finding was that SVMs significantly outperformed KNN classifiers in this context, leading the authors to conclude that KNN is less suitable for high-level semantic tasks in this domain.

In copy detection, the "Glocal" approach demonstrated impressive efficiency, reaching processing speeds of 1/20th of real-time on a single core and up to 1/35th of real-time on dual cores. While the "Balanced" profile performed well against benchmarks, the "No False Alarms" (NOFA) profile proved too strict, resulting in a high miss rate. The authors also noted that the 20D descriptor lacks scale invariance, which impacted performance on heavily resized video copies (e.g., Picture-in-Picture type 1 transformations).

Conclusion

The NII submission successfully demonstrated that "Glocal" signatures can balance speed and robustness for large-scale video analysis. While the method excels in efficiency and resilience to cropping, future work will focus on enhancing the descriptor's scale invariance and optimizing detection thresholds to improve the trade-off between false alarms and detection sensitivity.

Summary updated

BibTeX
@inproceedings{le2009national,
  author = {Duy-Dinh Le and Sébastien Poullot and Xiaomeng Wu and Michael Nett and Michael E. Houle and Shinichi Satoh and Michel Crucianu},
  title = {{National Institute of Informatics, Japan at TRECVID 2009}},
  booktitle = {TRECVID},
  year = {2009},
  url = {https://openreview.net/forum?id=BZZyXIn_0Y},
}
Figure