Research map / Architectures and efficiency / CNN design
Is Second-Order Information Helpful for Large-Scale Visual Recognition?
Peihua Li, Jiangtao Xie, Qilong Wang, Wangmeng Zuo
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 first-order max pooling with plain covariance pooling | AlexNet · ImageNet 2012 val (1-crop) | top-1 error (%): 41.52 -> 40.41 | Improved |
| add matrix power normalization (alpha=1/2) to covariance pooling (ablation) | AlexNet cov pooling · ImageNet 2012 val (1-crop) | top-1 error (%): 40.41 -> 38.51 | Improved |
| add M-Fro normalization instead of MPN to covariance pooling (ablation) | AlexNet cov pooling · ImageNet 2012 val (1-crop) | top-1 error (%): 40.41 -> 39.87 | Improved |
| add M-l2 normalization instead of MPN to covariance pooling (ablation) | AlexNet cov pooling · ImageNet 2012 val (1-crop) | top-1 error (%): 40.41 -> 39.65 | Improved |
| combine M-Fro with MPN (ablation) | MPN-COV AlexNet · ImageNet 2012 val (1-crop) | top-1 error (%): 38.51 -> 39.93 | Hurt |
| combine M-l2 with MPN (ablation) | MPN-COV AlexNet · ImageNet 2012 val (1-crop) | top-1 error (%): 38.51 -> 39.62 | Hurt |
| add element-wise power normalization after MPN (ablation) | MPN-COV AlexNet · ImageNet 2012 val (1-crop) | top-1 error (%): degraded by over 1% | Hurt |
| use warm initialization from max-pooling network instead of random init (ablation) | MPN-COV AlexNet · ImageNet 2012 val (1-crop) | top-1 error (%): 38.51 -> 37.35 | Improved |
| reduce exponent alpha from 1 to about 1/2 (ablation) | AlexNet cov pooling · ImageNet 2012 val (1-crop) | top-1 error (%): 40.41 -> 38.51 | Improved |
| increase alpha above 1 (ablation) | AlexNet cov pooling · ImageNet 2012 val (1-crop) | top-1 error (%) | Hurt |
| use matrix power normalization instead of matrix logarithm (DeepO2P) | second-order pooling AlexNet · ImageNet 2012 val (1-crop) | top-1 error (%): 42.16 -> 38.51 | Improved |
| use MPN-COV instead of B-CNN element-wise power normalization | second-order pooling AlexNet · ImageNet 2012 val (1-crop) | top-1 error (%): 39.89 -> 38.51 | Improved |
Closest papers by meaning
- Contextual Convolutional Networks (ICLR 2023 · similarity 0.90)
- Bilinear Convolutional Neural Networks for Fine-Grained Visual Recognition (TPAMI 2017 · similarity 0.90)
- Conv2Former: A Simple Transformer-Style ConvNet for Visual Recognition (TPAMI 2024 · similarity 0.90)
- ObjectNet: A large-scale bias-controlled dataset for pushing the limits of object recognition models (NeurIPS 2019 · similarity 0.90)
- Involution: Inverting the Inherence of Convolution for Visual Recognition (CVPR 2021 · similarity 0.89)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 2024 · similarity 0.88)
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.
