Research map / Vision-language and multimodal / CLIP and zero-shot
Training-free Discriminative Patch Mining for Robust Few-Shot Recognition with CLIP
Zhenzhang Ye, Duolikun Danier, Bo Zhao, Hakan Bilen
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 preliminary filtering (ablation) | CDPS hybrid classifier · UCF101 (4-shot) | accuracy (%): 71.05 -> 77.24 | Improved |
| add preliminary filtering (ablation) | CDPS hybrid classifier · EuroSAT (4-shot) | accuracy (%): 77.80 -> 80.89 | Improved |
| add preliminary filtering (ablation) | CDPS hybrid classifier · FGVCAircraft (4-shot) | accuracy (%): 38.47 -> 39.84 | Improved |
| add adaptive lambda combination instead of static 0.5 (ablation) | CDPS hybrid classifier · FGVCAircraft (4-shot) | accuracy (%): 32.52 -> 39.84 | Improved |
| add adaptive lambda combination instead of static 0.5 (ablation) | CDPS hybrid classifier · StanfordCars (4-shot) | accuracy (%): 86.87 -> 88.46 | Improved |
| add adaptive lambda combination instead of static 0.5 (ablation) | CDPS hybrid classifier · UCF101 (4-shot) | accuracy (%): 75.95 -> 77.24 | Improved |
| add the inter-class ambiguity term to the patch score (ablation) | CDPS · DTD, Aircraft, Pets, Birds (4-shot) | accuracy drop when removed: 3.07 average drop without inter | Improved |
| add the intra-class consistency term to the patch score (ablation) | CDPS · DTD, Aircraft, Pets, Birds (4-shot) | accuracy drop when removed: 4.67 average drop without intra | Improved |
| use leave-one-out selected lambda instead of empirically optimal lambda (ablation) | CDPS · 12 datasets | accuracy (%): 0.47% average accuracy gap (0.81 at 2 shots, 0.32 at 16) | No clear effect |
| increase number of discriminative patches K to 30 (ablation) | CDPS · FGVCAircraft | accuracy | No clear effect |
| fuse features from layers 2, 6 and 12 via linear combination | CDPS · ImageNet-V2 | accuracy: +1.32% | Improved |
| replace training-free baselines with CDPS in 8-shot ImageNet runtime | CDPS vs TIMO · ImageNet | time (s): 305.7 -> 80.6 | Improved |
Closest papers by meaning
- A Hard-to-Beat Baseline for Training-free CLIP-based Adaptation (ICLR 2024 · similarity 0.88)
- Text and Image Are Mutually Beneficial: Enhancing Training-Free Few-Shot Classification with CLIP (AAAI 2025 · similarity 0.88)
- Seeing What Matters: Empowering CLIP with Patch Generation-to-Selection (CVPR 2025 · similarity 0.87)
- Interpreting and Analysing CLIP's Zero-Shot Image Classification via Mutual Knowledge (NeurIPS 2024 · similarity 0.87)
- Large Language Models are Good Prompt Learners for Low-Shot Image Classification (CVPR 2024 · similarity 0.87)
- Rethinking Prior Information Generation with CLIP for Few-Shot Segmentation (CVPR 2024 · 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.
