Research map / Label-efficient and robust learning / Domain adaptation and generalization
The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation
Giacomo Zara, Alessandro Conti, Subhankar Roy, Stéphane Lathuilière, Paolo Rota, Elisa Ricci
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 |
|---|---|---|---|
| use zero-shot CLIP ViT-B/32 instead of best SFVUDA method | SFVUDA · Daily-DA | accuracy (%): 39.6 -> 48.1 | Improved |
| add unsupervised target adapter trained on CLIP pseudo-labels (ablation) | zero-shot CLIP · Daily-DA | accuracy (%): 48.1 -> 50.8 | Improved |
| ensemble CLIP, source and target adapters (ablation) | zero-shot CLIP · Daily-DA | accuracy (%): 48.1 -> 53.8 | Improved |
| ensemble CLIP, source and target adapters (ablation) | zero-shot CLIP · UCF-HMDB_full | accuracy (%): 89.7 -> 92.6 | Improved |
| use source adapter alone instead of CLIP (ablation) | zero-shot CLIP · Daily-DA | accuracy (%): 48.1 -> 41.5 | Hurt |
| increase number of text templates from 1 to 12 (ablation) | DALL-V inference · Daily-DA, UCF-HMDB_full | accuracy | Improved |
| increase templates beyond 12 (ablation) | DALL-V inference · Daily-DA, UCF-HMDB_full | accuracy | No clear effect |
| fine-tune full vision encoder instead of adapter in distillation (ablation) | DALL-V student · UCF-HMDB_full | accuracy (%): 91.0 -> 91.8 | Mixed |
| use original CLIP prompts instead of short action-oriented prompts (ablation) | zero-shot CLIP pseudo-labelling · UCF-HMDB_full | accuracy (%): 89.7 -> 83.2 | Hurt |
| replace 'video' with 'image' in prompts (ablation) | zero-shot CLIP · UCF-HMDB_full | accuracy (%): 89.7 -> 89.8 | No clear effect |
| DALL-V versus EXTERN on Sports-DA | SFVUDA · Sports-DA | accuracy (%): 83.2 -> 82.3 | Hurt |
Closest papers by meaning
- Source-Free Video Domain Adaptation With Spatial-Temporal-Historical Consistency Learning (CVPR 2023 · similarity 0.91)
- Source-Free Video Domain Adaptation by Learning Temporal Consistency for Action Recognition (ECCV 2022 · similarity 0.89)
- Source-Free Domain Adaptation with Frozen Multimodal Foundation Model (CVPR 2024 · similarity 0.89)
- Unsupervised Video Domain Adaptation with Masked Pre-Training and Collaborative Self-Training (CVPR 2024 · similarity 0.88)
- Domain-Agnostic Mutual Prompting for Unsupervised Domain Adaptation (CVPR 2024 · similarity 0.87)
- Modality-Collaborative Test-Time Adaptation for Action Recognition (CVPR 2024 · similarity 0.87)
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.
