Research map / Architectures and efficiency / CNN design
Analyzing Filters Toward Efficient ConvNet
Takumi Kobayashi
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 |
|---|---|---|---|
| Replace full convolution filters with orthonormal steerable bases from Analysis 1 | VGG-M · ILSVRC2014 validation | top-5 error rate (%): 12.96 -> 13.19 | Hurt |
| Use orthonormal steerable bases instead of random bases (ablation) | VGG-M · ILSVRC2014 validation | top-5 error rate (%): 13.65 -> 13.19 | Improved |
| Require orthogonality in steerable filter bases (ablation) | VGG-M · ILSVRC2014 validation | top-5 error rate (%): 13.64 -> 13.19 | Improved |
| Apply Analysis 1 convolution bases and dense-BoW max-pooling | AlexNet · ILSVRC2014 validation | top-5 error rate (%): 16.58 -> 14.08 | Improved |
| Apply Analysis 1 convolution bases and dense-BoW max-pooling | VGG-F · ILSVRC2014 validation | top-5 error rate (%): 16.36 -> 14.41 | Improved |
| Apply Analysis 1 convolution bases and dense-BoW average-pooling | VGG-S · ILSVRC2014 validation | top-5 error rate (%): 13.19 -> 11.57 | Improved |
| Replace standard FC layer with dense-BoW average-pooling features | VGG-vd-16 · ILSVRC2014 validation | top-5 error rate (%): 7.88 -> 7.10 | Improved |
| Replace standard FC layer with BoW features | VGG-vd-16 · MIT67 | classification accuracy (%): 75.66 -> 79.12 | Improved |
| Replace standard FC layer with BoW features | VGG-vd-16 · Bird200 | classification accuracy (%): 68.13 -> 77.60 | Improved |
| Replace conv5 and subsequent average-pooling with BoW (ablation) | ResNet-50 · ILSVRC2014 validation | top-5 error rate (%): 6.35 -> 6.23 | Improved |
| Apply Analysis 1 3×3 bases at conv2–4 (ablation) | ResNet-50 · ILSVRC2014 validation | top-5 error rate (%): 6.35 -> 6.42 | Hurt |
| Add Analysis 1 5×5 filters at conv4 (ablation) | ResNet-50 · ILSVRC2014 validation | top-5 error rate (%): 6.31 -> 6.01 | Improved |
Closest papers by meaning
- CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters (CVPR 2022 · similarity 0.88)
- Conv2Former: A Simple Transformer-Style ConvNet for Visual Recognition (TPAMI 2024 · similarity 0.88)
- Factors of Transferability for a Generic ConvNet Representation (TPAMI 2016 · similarity 0.88)
- Learning Versatile Convolution Filters for Efficient Visual Recognition (TPAMI 2021 · similarity 0.88)
- ConvNeXt V2: Co-Designing and Scaling ConvNets With Masked Autoencoders (CVPR 2023 · similarity 0.87)
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 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.
