Research map / Detection and segmentation / CNN features
Shape or Texture: Understanding Discriminative Features in CNNs
Read the paperCheck 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 |
|---|---|---|---|
| train ResNet50 solely on Stylized ImageNet instead of ImageNet (ablation) | ResNet50 stage-5 representation · Stylized PASCAL VOC dimensionality estimation | number of shape-encoding neurons |z_shape|: 349 -> 536 | Helped |
| train ResNet50 on Stylized ImageNet instead of ImageNet (shape bias) (ablation) | ResNet50 · Read-out binary/semantic segmentation on VOC12 val | Sem mIoU: 61.6 -> 53.7 | Hurt |
| train read-out on multi-stage features (f1 to f5) instead of last stage only (ablation) | ResNet50 IN-SEN · VOC12 val shape encoding | Bin and Sem mIoU: Bin/Sem improved (multi-stage vs f5 alone, e.g. 70.8/33.9 for f1 vs 70.6/50.9 for f5 alone) | Helped |
| biasing ResNet50 toward shape (SIN or SIN+IN) changes shape encoding mainly in the last two stages (ablation) | ResNet50 stage-wise shape dimensionality · Stage-wise dimensionality (Table 3) | percentage of shape neurons: f1-f3 essentially unchanged across training types; differences concentrated in f4-f5 | No clear effect |
| SIN-trained model evaluated on stylized VOC: large gap between binary and semantic mIoU | shape-biased ResNet50 (SIN) · Stylized VOC12 val | Bin vs Sem mIoU: Bin 60.3 vs Sem 26.4 | Mixed |
| remove top 300 shape-specific neurons from read-out inputs (ablation) | SIN-trained ResNet50 vs IN-trained ResNet50 · VOC12 val (Table 7) | Sem mIoU: SIN 53.7 -> 25.8 (drop 27.9); IN 50.9 -> 38.0 (drop 12.9) | Hurt |
| remove top 300 texture-specific neurons instead of shape neurons (ablation) | SIN-trained ResNet50 · VOC12 val (Table 7) | Sem mIoU: 53.7 -> 27.9 | Hurt |
| use a 3-layer instead of 1-layer read-out on IN-SEN features (ablation) | ResNet50 IN-SEN · VOC12 val, Table 4 | Sem mIoU: 50.9 -> 61.6 (ResNet50, 1 layer vs 3 layers as printed) | Helped |
Closest papers by meaning
- The Origins and Prevalence of Texture Bias in Convolutional Neural Networks (NeurIPS 2020 · similarity 0.95)
- ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness (ICLR 2019 · similarity 0.94)
- ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression (NeurIPS 2025 · similarity 0.92)
- Visual Anagrams Reveal Hidden Differences in Holistic Shape Processing Across Vision Models (NeurIPS 2025 · similarity 0.89)
- Hidden in Plain Sight: Evaluating Abstract Shape Recognition in Vision-Language Models (NeurIPS 2024 · similarity 0.87)
- The Shape of Events: Edge-Based Inductive Biases via Cross-Domain Distillation (NeurIPS 2026 · similarity 0.87)
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.