Research map / Vision-language and multimodal / Human and embodied modeling
Describing Textures using Natural Language
Chenyun Wu, Mikayla Timm, Subhransu Maji
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 BERT instead of mean-pooled word embeddings as phrase encoder (ablation) | MetricLearning · DTD2 validation phrase retrieval | MAP: 18.80 -> 31.68 | Improved |
| use BERT instead of biLSTM as phrase encoder (ablation) | MetricLearning · DTD2 validation phrase retrieval | MAP: 23.53 -> 31.68 | Improved |
| use BERT instead of ELMo as phrase encoder (ablation) | MetricLearning · DTD2 validation phrase retrieval | MAP: 28.13 -> 31.68 | Improved |
| use BERT instead of mean-pooled word embeddings as phrase encoder (ablation) | MetricLearning · DTD2 validation image retrieval from phrase | MAP: 7.19 -> 15.22 | Improved |
| train metric learning on descriptions instead of phrases | MetricLearning · DTD2 image retrieval from description | MRR: 12.92 -> 13.95 | Improved |
| metric learning with BERT instead of classifier | DTD2 baselines · DTD2 test phrase retrieval | MAP: 27.12 -> 31.77 | Improved |
| metric learning with BERT instead of classifier | DTD2 baselines · DTD2 test image retrieval from phrase | MAP: 14.75 -> 13.50 | Hurt |
| metric-learning top-5 phrases instead of Show-Attend-Tell generation | description generation · DTD2 description generation | CIDEr: 37.47 -> 47.34 | Improved |
| use ResNet layers 2 and 4 features instead of layer 1 only (ablation) | Classifier · DTD2 validation phrase retrieval | MAP: 13.10 -> 26.57 | Improved |
| use metric learning on description instead of phrase for Color+Pattern query | MetricLearning · Synthetic textures Color+Pattern | R-Precision: 41.73 -> 35.77 | Hurt |
| models queried with foreground color of pattern vs random chance | MetricLearning - phrase · Synthetic textures Foreground | R-Precision: 46.55 vs random 50.00 | No clear effect |
| combine CUB attributes with DTD2 texture attributes | logistic regression on CUB · CUB 200-way classification | accuracy | Improved |
Closest papers by meaning
- The Devil is in the Fine-Grained Details: Evaluating Open-Vocabulary Object Detectors for Fine-Grained Understanding (CVPR 2024 · similarity 0.85)
- Language-Informed Visual Concept Learning (ICLR 2024 · similarity 0.85)
- Not Only Text: Exploring Compositionality of Visual Representations in Vision-Language Models (CVPR 2025 · similarity 0.85)
- Textural or Textual: How Vision-Language Models Read Text in Images (ICML 2025 · similarity 0.85)
- ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression (NeurIPS 2025 · similarity 0.85)
- The Scene Language: Representing Scenes with Programs, Words, and Embeddings (CVPR 2025 · similarity 0.85)
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.
