270 likes | 289 Views
Video Summarization via D eterminantal P oint P rocesses (DPP). Boqing Gong University of Southern California Joint work with Wei-Lun Chao, Kristen Grauman, and Fei Sha. Background Basic idea of DPP Sequential DPP ( NIPS 2014) Large-margin training of DPP Conclusion. Background.
E N D
Video Summarization viaDeterminantal Point Processes (DPP) Boqing Gong University of Southern California Joint work with Wei-Lun Chao, Kristen Grauman, and Fei Sha
Background • Basic idea of DPP • Sequential DPP (NIPS 2014) • Large-margin training of DPP • Conclusion
Background • Motivation: Indispensable for fast video browsing and retrieval • Representation: • Key frames / segments extraction • Subset Selection problem
Background • Video summarization is hard: • Individual selected frame: Representativeness • Selected frames as a whole: Diversity • Naïve solution:Clustering • Competing !
Background • Clustering works?
Video summarization: an overview • Video summarization is hard: • What criteria lead to user perspective? • What kind of models: • Supervised learning ! • Diverse subset with representative items
Background • How to model subset selection problem? • Structured prediction, submodular functions • Determinantal Point Processes (DPPs) [Alex Kulesza and Ben Taskar, 2012]
Background • Basic idea of DPP • Sequential DPP (NIPS 2014) • Large-margin training of DPP • Conclusion
Basic idea of DPP • Idea: A point process based on matrix determinant. • Formulation:M discrete items (binary decision)
Basic idea of DPP • Why diverse? • Extreme cases:
Basic idea of DPP • Learning in DPP: • 11
Background • Basic idea of DPP • Sequential DPP (NIPS 2014) • Large-margin training of DPP • Conclusion
Sequential DPP • Motivation: • The temporal structure of video is missing • Proposed Idea: • Sequential DPP via Markov properties
… • …
Sequential DPP • Modeling the sequential structure: • Conditional DPP: still a DPP !
Sequential DPP • Parameterization:
Inference and Learning • Inference: • Allow brute-force search in small chunks • Optimization:
Sequential DPP • Experimental setting: • 3 datasets: OVP (50), Youtube (39), Kodak (18) • Fisher vectors + Saliency + Contextual features • Evaluation: Recall, Precision, and F1 score • Comparison: unsupervised methods & vanilla DPP
Sequential DPP • Experimental Results:
Sequential DPP • Experimental Results:
Sequential DPP • Experimental Results:
Background • Basic idea of DPP • Sequential DPP (NIPS 2014) • Large-margin training of DPP • Conclusion
Learning parameters in DPP • Maximum likehood estimation • Focuses on observed data only • Large-margin training • Maximizes margin between observed and undesired data • Discriminative learning • More flexible: incorporating evaluation metrics
Large-margin training of DPP • More discriminative and flexible
Conclusion • Supervised learning for video summarization • DPPs: modeling diversity subset selection • Video structure: Sequential DPP • Parameterization: Neural networks • Future work • Better inference algorithms • Models beyond DPP (submodular)