| <Poster Width="1734" Height="1156"> |
| <Panel left="17" right="186" width="413" height="308"> |
| <Text>Highlight</Text> |
| <Text>- Pose video summarization</Text> |
| <Text>as a supervised learning problem</Text> |
| <Text>for subset selection</Text> |
| <Text>- Propose sequential determinantal</Text> |
| <Text>point process (seqDPP) as the</Text> |
| <Text>underlying probabilistic model</Text> |
| <Text>- Evaluate on three video</Text> |
| <Text>summarization tasks and obtain</Text> |
| <Text>state-of-the-art performance</Text> |
| </Panel> |
|
|
| <Panel left="18" right="503" width="411" height="635"> |
| <Text>Introduction</Text> |
| <Text>Video summarization: pressing need</Text> |
| <Text>- 100 hours of new Youtube video per min</Text> |
| <Text>- 422,000 CCTV cameras in London 24/7</Text> |
| <Text>Summaries by three users</Text> |
| <Figure left="25" right="650" width="395" height="118" no="1" OriWidth="0" OriHeight="0 |
| " /> |
| <Text>Challenges</Text> |
| <Text>- Heterogeneous subjects/categories</Text> |
| <Text>- Various temporal changing rates</Text> |
| <Text>- Subjective, disparate, and noisy labels</Text> |
| <Text>Previous work</Text> |
| <Text>- Criteria: representativeness vs. diversity</Text> |
| <Text>- Largely unsupervised, frame clustering</Text> |
| <Text>- Require sophisticated handcrafting</Text> |
| <Text>Our main idea</Text> |
| <Text>- Supervised learning from human</Text> |
| <Text>supplied annotations</Text> |
| <Text>- Summarization as subset selection</Text> |
| <Text>- Modeling temporal cue & diversity</Text> |
| </Panel> |
|
|
| <Panel left="446" right="187" width="413" height="948"> |
| <Text>Approach</Text> |
| <Text>Sequential DPP (seqDPP)</Text> |
| <Text>1. Partition video into T disjoint segments</Text> |
| <Text>2. Introduce subset selection (of frames)</Text> |
| <Text>variable Yt for each segment</Text> |
| <Text>3. Condition Yt on Yt-1 = yt-1 by DPP</Text> |
| <Figure left="452" right="386" width="403" height="287" no="2" OriWidth="0" OriHeight="0 |
| " /> |
| <Text>Parameterization of DPP kernel</Text> |
| <Text>- Linear embedding (L):</Text> |
| <Text>- Neural networks (NN)</Text> |
| <Text>Inference</Text> |
| <Text>Learning via MLE</Text> |
| <Text>- through gradient descent</Text> |
| <Text>In contrast, bag DPPs:</Text> |
| <Text>Model permutable items (no temporal info)</Text> |
| <Text>Often use quality-diversity kernel (limited)</Text> |
| <Text>Inference NP hard</Text> |
| </Panel> |
|
|
| <Panel left="874" right="189" width="841" height="252"> |
| <Text>Generating target summaries</Text> |
| <Text>User study on inter-annotator agreement</Text> |
| <Text>- Data: 100 videos from Open Video Project and Youtube</Text> |
| <Text>- Annotation: 5 user summaries per video</Text> |
| <Text>- Observation: high inter-annotator agreement</Text> |
| <Text>Generate target summaries by greedy search</Text> |
| <Figure left="1449" right="226" width="263" height="210" no="3" OriWidth="0.182238" OriHeight="0.110071 |
| " /> |
| </Panel> |
|
|
| <Panel left="876" right="456" width="838" height="577"> |
| <Text>Experiments</Text> |
| <Text>Setup</Text> |
| <Text>- Data: OVP (50), Youtube (39), Kodak (18)</Text> |
| <Text>- Feature: Fisher vector, saliency, context</Text> |
| <Text>- Evaluation: Precision, Recall, F-score</Text> |
| <Text>- Comparison: bag DPP and previous</Text> |
| <Text>(unsupervised) DT, STIMO, VSUMM</Text> |
| <Text> Results on Youtube and Kodak</Text> |
| <Figure left="898" right="702" width="454" height="116" no="4" OriWidth="0.605536" OriHeight="0.0490196 |
| " /> |
| <Text> Results on OVP</Text> |
| <Figure left="1397" right="540" width="293" height="282" no="5" OriWidth="0.287774" OriHeight="0.189394 |
| " /> |
| <Figure left="887" right="830" width="816" height="198" no="6" OriWidth="0.609573" OriHeight="0.104724 |
| " /> |
| </Panel> |
|
|
| <Panel left="876" right="1049" width="839" height="87"> |
| <Text>[1] S. Avila, A. Lopes, A. Luz Jr, A. Araujo. “VSUMM: A mechanism designed to produce static video summaries and</Text> |
| <Text>a novel evaluation method”. Pattern Recognition Letters, 32(1):56–68, 2011.</Text> |
| <Text>[2] A. Kulesza and B. Taskar. “Determinantal point processes for machine learning”. Foundations and Trends® in</Text> |
| <Text>Machine Learning, 5(2-3):123–286, 2012.</Text> |
| </Panel> |
|
|
| </Poster> |
|
|
