Research map / Architectures and efficiency / CNN design
Dynamic Convolutions: Exploiting Spatial Sparsity for Faster Inference
Check 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 |
|---|---|---|---|
| Use dynamic convolutions with θ=0.25 instead of dense MobileNetV2 x0.75 | MobileNetV2 · Food-101 test | images/sec: 508 -> 724 | Helped |
| Use the binary Gumbel-Softmax reformulation instead of standard Gumbel-Softmax (ablation) | DynConv MobileNetV2, θ=0.25 · Food-101 test | images/sec: 522 -> 724 | Helped |
| Use dynamic convolutions with θ=0.25 instead of dense ShuffleNetV2 | ShuffleNetV2 · Food-101 test | images/sec: 710 -> 781 | Helped |
| Add a 1×1 convolution mask unit to DynConv ShuffleNetV2 (ablation) | DynConv ShuffleNetV2, θ=0.25 · Food-101 test | images/sec: 781 -> 889 | Helped |
| Use dynamic convolutions with θ=0.75 instead of the dense baseline | 4-stack hourglass · MPII validation | images/sec: 30 -> 33 | Helped |
| Use dynamic convolutions with θ=0.50 instead of the dense baseline | 4-stack hourglass · MPII validation | images/sec: 30 -> 48 | Helped |
| Use dynamic convolutions with θ=0.25 instead of the dense baseline | 4-stack hourglass · MPII validation | PCKh@0.5: 88.1 -> 87.5 | Hurt |
| Use dynamic convolutions with θ=0.125 instead of the dense baseline | 4-stack hourglass · MPII validation | images/sec: 30 -> 85 | Helped |
| Use a squeeze mask unit instead of a 1×1 convolution mask unit (ablation) | DynConv 4-stack hourglass, θ=0.125 · MPII validation | PCKh@0.5: 86.7 -> 87.0 | Helped |
| Use a squeeze mask unit instead of a 1×1 convolution mask unit (ablation) | DynConv 4-stack hourglass, θ=0.125 · MPII validation | images/sec: 85 -> 76 | Hurt |
| Add upper and lower bound constraints to the network-wise sparsity loss (ablation) | DynConv 4-stack hourglass, θ=0.125 · MPII validation | PCKh@0.5: 86.1 -> 86.7 | Helped |
| Use the proposed upper/lower-bound sparsity criterion instead of a per-layer budget (ablation) | DynConv 4-stack hourglass, θ=0.125 · MPII validation | PCKh@0.5: 86.3 -> 86.7 | Helped |
Closest papers by meaning
- Convolutional Pose Machines (CVPR 2016 · similarity 0.87)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.87)
- Glance and Focus: a Dynamic Approach to Reducing Spatial Redundancy in Image Classification (NeurIPS 2020 · similarity 0.87)
- Temporal Residual Networks for Dynamic Scene Recognition (CVPR 2017 · similarity 0.87)
- Dynamic Convolution: Attention Over Convolution Kernels (CVPR 2020 · similarity 0.86)
- Efficient Deformable ConvNets: Rethinking Dynamic and Sparse Operator for Vision Applications (CVPR 2024 · 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.