Research map / Detection and segmentation / CNN features
Mind the Pad -- CNNs Can Develop Blind Spots
Bilal Alsallakh, Narine Kokhlikyan, Vivek Miglani, Jun Yuan, Orion Reblitz-Richardson
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 |
|---|---|---|---|
| Train with 225 × 225 rather than 224 × 224 input (ablation) | MobileNet · ImageNet | Top-1 accuracy: 68.19 -> 68.80 | Improved |
| Train with 225 × 225 rather than 224 × 224 input (ablation) | ResNet-18 · ImageNet | Top-1 accuracy: 69.93 -> 70.27 | Improved |
| Train with 225 × 225 rather than 224 × 224 input (ablation) | ResNet-34 · ImageNet | Top-1 accuracy: 73.30 -> 73.72 | Improved |
| Train with 225 × 225 rather than 224 × 224 input (ablation) | ResNet-50 · ImageNet | Top-1 accuracy: 75.65 -> 76.01 | Improved |
| Train with 225 × 225 rather than 224 × 224 input (ablation) | ResNet-101 · ImageNet | Top-1 accuracy: 77.37 -> 77.67 | Improved |
| Replace Zero Padding with Mirror Padding (ablation) | SSD with MobileNet-v1 · BSTLD test | AP@.20IOU: 80.24% -> 83.20% | Improved |
| Replace Zero Padding with Mirror Padding (ablation) | SSD with MobileNet-v1 · BSTLD test | AP@.50IOU: 49.58% -> 57% | Improved |
| Replace Zero Padding with Mirror Padding (ablation) | SSD with MobileNet-v1 · BSTLD test | AP@.75IOU: 3.7% -> 8.44% | Improved |
| Replace Zero Padding with Mirror Padding (ablation) | SSD with MobileNet-v1 · BSTLD test | AP@.90IOU: 0.007% -> 0.02% | Improved |
Closest papers by meaning
- ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression (NeurIPS 2025 · similarity 0.84)
- Partial Convolution for Padding, Inpainting, and Image Synthesis (TPAMI 2022 · similarity 0.84)
- Convolution with even-sized kernels and symmetric padding (NeurIPS 2019 · similarity 0.84)
- Emergence of Shape Bias in Convolutional Neural Networks through Activation Sparsity (NeurIPS 2023 · similarity 0.84)
- On Translation Invariance in CNNs: Convolutional Layers Can Exploit Absolute Spatial Location (CVPR 2020 · similarity 0.84)
- Inductive Bias of Deep Convolutional Networks through Pooling Geometry (ICLR 2017 · similarity 0.84)
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.
