Research map / Vision-language and multimodal / CLIP and zero-shot
Think Twice: Test-Time Reasoning for Robust CLIP Zero-Shot Classification
Shenyu Lu, Zhaoying Pan, Xiaoqian Wang
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 |
|---|---|---|---|
| remove spurious backgrounds (core object only, black background) | CLIP ViT-L/14 zero-shot · Waterbirds | worst-group accuracy: 34.13 -> 91.39 | Improved |
| add PCA then K-means localization of the core object before classification (ablation) | CLIP ViT-L/14 zero-shot · Waterbirds | worst-group accuracy: 47.36 -> 74.14 | Improved |
| use UMAP for dimensionality reduction instead of PCA+K-means (ablation) | CLIP ViT-L/14 zero-shot · Waterbirds | worst-group accuracy: 47.36 -> 36.01 | Hurt |
| add SAM masks to Grounding DINO boxes (ablation) | CLIP ViT-L/14 zero-shot · Waterbirds | worst-group accuracy: 66.80 -> 87.79 | Improved |
| reduce to 3 dimensions with PCA before K-means instead of clustering 768-d features (ablation) | TTR · Waterbirds | worst-group accuracy: 69.43 -> 74.14 | Improved |
| add TTR to zero-shot CLIP on general datasets | CLIP ViT-L/14 · General datasets average | worst-group accuracy: 54.56 -> 70.58 | Improved |
| add TTR on UrbanCars | CLIP ViT-L/14 · UrbanCars | worst-group accuracy: 8.00 -> 34.00 | Improved |
| add TTR on Metashift | CLIP ViT-L/14 · Metashift | worst-group accuracy: 89.54 -> 90.85 | No clear effect |
| use a queue of historical samples for PCA (ablation) | TTR · Waterbirds | worst-group accuracy | Improved |
| increase number of K-means clusters beyond 3 (ablation) | TTR · Waterbirds | worst-group accuracy | Hurt |
| fill masked regions with zeros instead of mean or Gaussian blur (ablation) | TTR · Waterbirds | worst-group accuracy | Improved |
| apply TTR to ViT-B/32 and ViT-H/14 (ablation) | CLIP · Spurious-correlation benchmarks | worst-group accuracy | Improved |
Closest papers by meaning
- Interpreting and Analysing CLIP's Zero-Shot Image Classification via Mutual Knowledge (NeurIPS 2024 · similarity 0.93)
- PerceptionCLIP: Visual Classification by Inferring and Conditioning on Contexts (ICLR 2024 · similarity 0.90)
- Transductive Zero-Shot and Few-Shot CLIP (CVPR 2024 · similarity 0.90)
- A Hard-to-Beat Baseline for Training-free CLIP-based Adaptation (ICLR 2024 · similarity 0.89)
- TPR: Topology-Preserving Reservoirs for Generalized Zero-Shot Learning (NeurIPS 2024 · similarity 0.89)
- Exploring Regional Clues in CLIP for Zero-Shot Semantic Segmentation (CVPR 2024 · similarity 0.89)
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.
