Research map / Vision-language and multimodal / CLIP and zero-shot
SOTA: Self-adaptive Optimal Transport for Zero-Shot Classification with Multiple Foundation Models
Zhanxuan Hu, Qiyu Xu, Yu Duan, Yonghang Tai, Huafeng Li
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 |
|---|---|---|---|
| add vision-only foundation models (DINO) to the VLM ensemble (ablation) | SOTA · natural, remote sensing, medical (ablation average) | top-1 accuracy (%): +11.1 / +8.6 / +12.9 over base (full SOTA) | Improved |
| integrate multiple VLMs only (Only-P-hat), no VFM (ablation) | SOTA · natural, remote sensing, medical (ablation average) | top-1 accuracy (%): +0.2 / +0.3 / +0.8 | Improved |
| use self-adaptive weights instead of fixed weights (ablation) | SOTA · natural, remote sensing, medical (ablation average) | top-1 accuracy (%): modest improvement | Improved |
| use self-adaptive weights instead of non-adaptive weights with CLIP and MUSK (ablation) | SOTA · 5 medical datasets (CLIP+MUSK) | average top-1 accuracy (%): 71.8 -> 73.1 | Improved |
| use joint learning of GMM and transport plan instead of disjoint learning (ablation) | SOTA · natural, remote sensing, medical (ablation average) | top-1 accuracy (%) | Improved |
| swap CLIP ViT-B/16 for a stronger visual encoder (DINOv3) in the ensemble | SOTA · 11 natural datasets, transductive | average top-1 accuracy (%): 72.5 -> 77.4 | Improved |
| add DINOv2 and DINOv3 jointly versus DINOv3 alone | SOTA · 11 natural datasets, transductive | average top-1 accuracy (%): 77.4 -> 77.8 | No clear effect |
| ensemble two CLIP variants (ViT-B/16 + RN50) on Aircraft | SOTA · Aircraft, transductive | accuracy (%): 24.7 -> 22.6 | Hurt |
| ensemble four pathology/general VLMs with SOTA instead of applying TransCLIP to the best model | SOTA · 5 medical pathology datasets, transductive | average top-1 accuracy (%): 76.0 -> 83.9 | Improved |
| ensemble four remote sensing VLMs with SOTA instead of applying TransCLIP to the best model | SOTA · 10 remote sensing datasets, transductive | average top-1 accuracy (%): 76.2 -> 81.5 | Improved |
| use SOTA in inductive mode versus test-time adaptation methods | SOTA · 11 natural datasets, inductive | average top-1 accuracy (%): 73.3 -> 77.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ADAPT — ADAPT: Vision-Language Navigation With Modality-Aligned Action Prompts (CVPR 2022) · better on 1 of 1 shared result
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
- S3A: Towards Realistic Zero-Shot Classification via Self Structural Semantic Alignment (AAAI 2024 · 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 0.87)
- A Closer Look at the Few-Shot Adaptation of Large Vision-Language Models (CVPR 2024 · similarity 0.87)
- AWT: Transferring Vision-Language Models via Augmentation, Weighting, and Transportation (NeurIPS 2024 · similarity 0.87)
- Conformal Prediction for Zero-Shot Models (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.
