Research map / Video / Video-language understanding
Bi-calibration Networks for Weakly-Supervised Video Representation Learning
Fuchen Long, Ting Yao, Zhaofan Qiu, Xinmei Tian, Jiebo Luo, Tao Mei
Read the paperCheck 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 |
|---|---|---|---|
| cluster titles into a text vocabulary and use multi-label classification (TS) instead of regression (RG) (ablation) | text supervision · K400 linear, YOVO-3M | top-1 accuracy (%): 64.5 -> 72.5 | Improved |
| use TS instead of triplet ranking (TR) (ablation) | text supervision · K400 linear, YOVO-3M | top-1 accuracy (%): 67.2 -> 72.5 | Improved |
| use TS instead of the variant TS- with single-cluster assignment (ablation) | text supervision · K400 linear, YOVO-3M | top-1 accuracy (%): 71.5 -> 72.5 | Improved |
| use TS instead of InfoNCE contrastive learning (ablation) | video-text learning · K400 linear | top-1 accuracy (%): 72.2 -> 72.5 | Improved |
| add text supervision to query supervision (ablation) | QS · K400 linear, YOVO-3M | top-1 accuracy (%): 71.1 -> 73.1 | Improved |
| add text-to-query calibration (ablation) | QS+TS · K400 linear, YOVO-3M | top-1 accuracy (%): 73.1 -> 73.8 | Improved |
| add query-to-text calibration (ablation) | QS+TS · K400 linear, YOVO-3M | top-1 accuracy (%): 73.1 -> 73.6 | Improved |
| use both calibration directions (ablation) | QS+TS · K400 linear, YOVO-3M | top-1 accuracy (%): 73.1 -> 74.2 | Improved |
| add bi-calibration on top of QS+TS (ablation) | QS+TS · UCF101 linear, YOVO-3M | top-1 accuracy (%): 92.3 -> 93.5 | Improved |
| add bi-calibration on top of QS+TS (ablation) | QS+TS · HMDB51 linear, YOVO-3M | top-1 accuracy (%): 65.9 -> 67.6 | Improved |
| pre-train with BCN instead of CPD | fine-tuned ResNet-50 · K400 fine-tune, YOVO-3M | top-1 accuracy (%): 73.9 -> 78.5 | Improved |
| scale pre-training data from YOVO-3M to YOVO-10M | BCN · K400 fine-tune | top-1 accuracy (%): 78.5 -> 79.1 | Improved |
Closest papers by meaning
- Weakly Supervised Video Representation Learning With Unaligned Text for Sequential Videos (CVPR 2023 · similarity 0.87)
- Weakly Supervised Summarization of Web Videos (ICCV 2017 · similarity 0.86)
- WINNER: Weakly-Supervised hIerarchical decompositioN and aligNment for Spatio-tEmporal Video gRounding (CVPR 2023 · similarity 0.86)
- Learning Multi-Scale Video-Text Correspondence for Weakly Supervised Temporal Article Gronding (AAAI 2024 · similarity 0.85)
- TOPA: Extending Large Language Models for Video Understanding via Text-Only Pre-Alignment (NeurIPS 2024 · similarity 0.85)
- Mitigating Semantic Collapse in Partially Relevant Video Retrieval (NeurIPS 2025 · similarity 0.85)
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.
