Research map / Video / Video-language understanding
Large Scale Video Representation Learning via Relational Graph Clustering
Hyodong Lee, Joonseok Lee, Joe Yue-Hei Ng, Paul Natsev
Check your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace random negatives with graph-cluster smart negatives | CDML-style triplet learning · Related video retrieval | MAP: 2.85% -> 3.57% | Improved |
| Replace random negatives with graph-cluster smart negatives | CDML-style triplet learning · Related video retrieval | NDCG@60: 6.05% -> 7.19% | Improved |
| Use harder negatives sampled from lower-level clusters (ablation) | GCML · Related video retrieval | MAP: 3.37% -> 3.43% -> 3.57% | Improved |
| Use harder negatives sampled from lower-level clusters (ablation) | GCML · Related video retrieval | NDCG@60: 6.64% -> 6.91% -> 7.19% | Improved |
| Increase GCML training triplets from 45M to 420M | GCML · Related video retrieval | MAP: 3.57% -> 3.74% | Improved |
| Train with graph-cluster pseudo-label classification instead of triplet loss | Video embedding learning · Related video retrieval | MAP: 3.41% vs 3.34% (CDML 458M) | Mixed |
| Limit sampled examples to 20–100 rather than 1,000 per cluster (ablation) | CLC · Related video retrieval | MAP: 3.24% -> 3.42% | Improved |
| Increase sampled classes from 10,000 to 200,000 (ablation) | CLC · Related video retrieval | MAP: 3.21% -> 3.42% | Improved |
| Increase sampled classes from 200,000 to 400,000 (ablation) | CLC · Related video retrieval | MAP: 3.42% -> 3.42% | No clear effect |
| Use GCML instead of CDML with small training sets | Video annotation · YouTube-8M | MAP: 51.7% -> 52.5% | Improved |
| Use GCML instead of CDML with large training sets | Video annotation · YouTube-8M | MAP: 55.5% vs 55.5% | No clear effect |
| Use GCML embeddings instead of CDML embeddings | Video annotation · Sports-1M | Hit@1 (embedding only): 67.0% -> 68.7% | Improved |
Closest papers by meaning
- GOCA: Guided Online Cluster Assignment for Self-Supervised Video Representation Learning (ECCV 2022 · similarity 0.87)
- Multiview Pseudo-Labeling for Semi-Supervised Learning From Video (ICCV 2021 · similarity 0.85)
- Evolving Losses for Unsupervised Video Representation Learning (CVPR 2020 · similarity 0.85)
- GraphVid: It Only Takes a Few Nodes to Understand a Video (ECCV 2022 · similarity 0.85)
- Enhancing Self-Supervised Video Representation Learning via Multi-Level Feature Optimization (ICCV 2021 · similarity 0.84)
- Hierarchical Correlation Clustering and Tree Preserving Embedding (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
