Research map / Vision-language and multimodal / CLIP and zero-shot
Let's Split Up: Zero-Shot Classifier Edits for Fine-Grained Video Understanding
Kaiting Liu, Hazel Doughty
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 modifier retrieval on classifier weights instead of a VLM text-similarity baseline (ablation) | zero-shot category splitting · SSv2-Split subset A | generality: 27.6 -> 45.0 | Improved |
| add modifier alignment on top of modifier retrieval (ablation) | zero-shot category splitting · SSv2-Split subset A | generality: 45.0 -> 46.3 | Improved |
| finetune the full backbone with one example per subcategory (ablation) | one-shot finetuning · SSv2-Split subset A | locality: 98.4 -> 0.0 | Hurt |
| restrict finetuning to the new head only (ablation) | one-shot finetuning · SSv2-Split subset A | locality: 0.0 -> 98.4 | Improved |
| initialize new head with modifier alignment instead of coarse weight (ablation) | one-shot finetuning · SSv2-Split subset A | generality: 48.4 -> 52.8 | Improved |
| use one-shot head finetuning instead of full-data finetuning (ablation) | category splitting · SSv2-Split subset A | mean of generality and locality: 54.4 -> 75.5 | Improved |
| use stronger pretraining (MVD) instead of training from scratch (ablation) | modifier alignment · SSv2-Split subset A | generality: 37.0 -> 46.3 | Improved |
| use video-text text encoders (VideoPrism) instead of CLIP (ablation) | modifier alignment · SSv2-Split subset A | generality: 46.3 -> 46.5 | No clear effect |
| use InternVideo2 text encoder instead of CLIP (ablation) | modifier alignment · SSv2-Split subset A | generality: 46.3 -> 36.9 | Hurt |
| scale backbone from ViT-Small to ViT-Base/Large (ablation) | modifier alignment · SSv2-Split subset A | generality: 46.3 -> 43.6 (Base), 45.0 (Large) | No clear effect |
| add category and pseudo-coarse pairs to alignment training (ablation) | modifier alignment · FineGym-Split subset A | generality: 27.6 -> 34.2 | Improved |
| add category and pseudo-coarse pairs to alignment training (ablation) | modifier alignment · SSv2-Split subset B | generality: 40.2 -> 38.4 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- FG-CLIP — FG-CLIP: Fine-Grained Visual and Textual Alignment (ICML 2025) · better on 4 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Revisiting Classifier: Transferring Vision-Language Models for Video Recognition (AAAI 2023 · similarity 0.89)
- Rethinking Zero-Shot Learning: A Conditional Visual Classification Perspective (ICCV 2019 · similarity 0.89)
- Rethinking Zero-Shot Video Classification: End-to-End Training for Realistic Applications (CVPR 2020 · similarity 0.89)
- Learning to Model Relationships for Zero-Shot Video Classification (TPAMI 2020 · similarity 0.88)
- Improved Zero-Shot Classification by Adapting VLMs with Text Descriptions (CVPR 2024 · similarity 0.87)
- Rethinking Few-Shot Adaptation of Vision-Language Models in Two Stages (CVPR 2025 · similarity 0.86)
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.
