Research map / Architectures and efficiency / CNN design
Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks With Octave Convolution
Yunpeng Chen, Haoqi Fan, Bing Xu, Zhicheng Yan, Yannis Kalantidis, Marcus Rohrbach, Shuicheng Yan, Jiashi Feng
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 convolutions with OctConv at alpha=0.125 (ablation) | ResNet-50 · ImageNet val | top-1 accuracy: 77.0 -> 78.2 | Improved |
| raise low-frequency ratio alpha to 0.75 (ablation) | OctConv ResNet-50 · ImageNet val | top-1 accuracy: 77.0 -> 76.7 (4.1 -> 1.9 GFLOPs) | Mixed |
| add Low->High information path (ablation) | OctConv ResNet-50 alpha 0.5 · ImageNet val | top-1 accuracy: 76.0 -> 76.4 | Improved |
| add High->Low information path (ablation) | OctConv ResNet-50 alpha 0.5 · ImageNet val | top-1 accuracy: 76.0 -> 76.4 | Improved |
| use average pooling instead of strided convolution for down-sampling (ablation) | OctConv ResNet-50 alpha 0.5 · ImageNet val | top-1 accuracy: 76.3 -> 77.4 | Improved |
| use average pooling instead of max pooling for down-sampling (ablation) | OctConv ResNet-50 alpha 0.5 · ImageNet val | top-1 accuracy: 77.0 -> 77.4 | Improved |
| larger test resolution (256 to 740) widens OctConv gain over baseline (ablation) | OctConv ResNet-50 alpha 0.5 · ImageNet val | top-1 accuracy: +0.7 -> +1.2 | Improved |
| apply OctConv to shallow ResNet-26 (alpha 0.25) (ablation) | ResNet-26 · ImageNet val | top-1 accuracy: 73.2 -> 76.1 | Improved |
| use OctConv with wider model at same FLOPs | MobileNet v1 · ImageNet val | top-1 accuracy: 70.3 -> 72.5 | Improved |
| apply OctConv alpha=0.1 | I3D trained from scratch · Kinetics-400 | top-1 accuracy: 72.6 -> 73.6 | Improved |
| apply OctConv alpha=0.5 (ablation) | I3D trained from scratch · Kinetics-400 | top-1 accuracy: 72.6 -> 72.1 | Hurt |
| apply OctConv alpha=0.1 | C2D ImageNet pretrained · Kinetics-400 | top-1 accuracy: 71.9 -> 73.8 | Improved |
Closest papers by meaning
- Glance and Focus: a Dynamic Approach to Reducing Spatial Redundancy in Image Classification (NeurIPS 2020 · similarity 0.87)
- Inference, Learning and Attention Mechanisms that Exploit and Preserve Sparsity in CNNs (IJCV 2020 · similarity 0.87)
- Deep Pyramidal Residual Networks (CVPR 2017 · similarity 0.87)
- Learning in the Frequency Domain (CVPR 2020 · similarity 0.86)
- Asymmetric Convolution: An Efficient and Generalized Method to Fuse Feature Maps in Multiple Vision Tasks (TPAMI 2024 · similarity 0.86)
- Multiscale Low-Frequency Memory Network for Improved Feature Extraction in Convolutional Neural Networks (AAAI 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.
