Research map / Image generation / Autoregressive and latent generation
Visualizing and Understanding Deep Texture Representations
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| replace Fisher Vector pooling with bilinear pooling | VGG-16 texture and scene descriptors · DTD, FMD, KTH-T2b, MIT indoor | mean per-class accuracy: 73.6 vs 72.9; 80.8 vs 81.6; 77.9 vs 77.9; 78.5 vs 79.0 | No clear effect |
| combine multiple input scales (ablation) | B-CNN · DTD, FMD, KTH-T2b, MIT indoor | mean per-class accuracy: 69.6 -> 72.9; 77.8 -> 81.6; 75.1 -> 77.9; 72.8 -> 79.0 | Helped |
| use higher CNN layers for bilinear features (ablation) | B-CNN · DTD, FMD, KTH-T2b, MIT indoor | classification accuracy: DTD: 42.9 -> 69.9; FMD: 49.6 -> 80.2; KTH-T2b: 59.9 -> 79.0; MIT indoor: 32.2 -> 72.8 | Helped |
| combine features from multiple layers (ablation) | B-CNN · DTD | classification accuracy: 69.9% -> 70.7% | Helped |
| combine features from multiple layers (ablation) | B-CNN · MIT indoor | classification accuracy: 72.8% -> 74.9% | Helped |
| fine-tune the network using the B-CNN architecture (ablation) | B-CNN · MIT indoor | classification accuracy: 72.8% -> 73.8% | Helped |
| fine-tune on the texture datasets (ablation) | B-CNN · texture recognition datasets | classification accuracy | No clear effect |
| use B-CNN instead of standard fully connected layers with spatial jittering | VGG-M · ImageNet LSVRC 2012 | top1 validation error: 39.6% -> 38.7% | Helped |
| increase B-CNN training augmentation from f1 to f25 (ablation) | B-CNN · ImageNet LSVRC 2012 | top1 validation error: 38.7% -> 37.1% | Helped |
| increase B-CNN augmentation from f5 to f25 (ablation) | B-CNN · ImageNet LSVRC 2012 | top1 validation error: <1% | No clear effect |
| initialize gradient-based texture synthesis with image quilting instead of a random image (ablation) | bilinear-feature texture synthesis · texture synthesis | optimization convergence: 250 iterations | Helped |
Closest papers by meaning
- The Origins and Prevalence of Texture Bias in Convolutional Neural Networks (NeurIPS 2020 · similarity 0.89)
- Texture Interpolation for Probing Visual Perception (NeurIPS 2020 · similarity 0.89)
- Shape or Texture: Understanding Discriminative Features in CNNs (ICLR 2021 · similarity 0.89)
- Deep Structure-Revealed Network for Texture Recognition (CVPR 2020 · similarity 0.88)
- ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression (NeurIPS 2025 · similarity 0.87)
- Visual Anagrams Reveal Hidden Differences in Holistic Shape Processing Across Vision Models (NeurIPS 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.