Research map / Architectures and efficiency / CNN design
Designing by Training: Acceleration Neural Network for Fast High-Dimensional Convolution
Longquan Dai, Liang Tang, Yuan Xie, Jinhui Tang
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 |
|---|---|---|---|
| use AccNet-designed acceleration on bilateral grid instead of handmade BG | bilateral grid acceleration · 2D Gaussian filtering sigma=2 | MSE: 0.952 -> 0.309 | Improved |
| use AccNet-designed acceleration on bilateral grid instead of handmade BG | bilateral grid acceleration · 5D Gaussian filtering sigma=2 | MSE: 1.804 -> 0.853 | Improved |
| use AccNet-designed acceleration on permutohedral lattice instead of handmade PL | permutohedral lattice acceleration · 2D Gaussian filtering sigma=2 | MSE: 0.541 -> 0.273 | Improved |
| use AccNet-designed acceleration on permutohedral lattice instead of handmade PL | permutohedral lattice acceleration · 5D Gaussian filtering sigma=2 | MSE: 1.712 -> 0.528 | Improved |
| replace three-layer CNN (cascaded 1-D convolutions) with AccNet (ablation) | acceleration network · 5D bilateral grid, bandwidth 5 | filtering error (MSE): 0.473 -> 0.461 | Improved |
| replace three-layer CNN with AccNet (ablation) | acceleration network · 2D bilateral grid, bandwidth 5 | filtering error (MSE): 0.245 -> 0.239 | Improved |
| replace three-layer CNN with AccNet (ablation) | acceleration network · 5D bilateral grid, bandwidth 5 | training time: 14h -> 7.6h | Improved |
| replace three-layer CNN with AccNet (ablation) | acceleration network · 2D bilateral grid, bandwidth 5 | training time: 12.5h -> 7.2h | Improved |
| use non-Gaussian (smooth exponential family) prior in bilateral solver via AccNet instead of Barron and Poole Gaussian | bilateral solver for stereo post-processing · Stereo matching All | RMS: 8.44 -> 7.79 | Improved |
| same, non-Gaussian bilateral solver | bilateral solver for stereo post-processing · Stereo matching All | MAE: 2.81 -> 2.13 | Improved |
| same, non-Gaussian bilateral solver | bilateral solver for stereo post-processing · Stereo matching NoOcc | bad 1%: 11.33 -> 10.41 | Mixed |
| accelerate Gaussian mixture kernels in fully connected CRF directly with AccNet | fully connected CRF inference · CRF segmentation | running time: save 60% of the time | Improved |
Closest papers by meaning
- Accelerating Very Deep Convolutional Networks for Classification and Detection (TPAMI 2016 · similarity 0.86)
- ESPACE: Accelerating Convolutional Neural Networks via Eliminating Spatial and Channel Redundancy (AAAI 2017 · similarity 0.86)
- Accelerating Sparse Convolution with Column Vector-Wise Sparsity (NeurIPS 2022 · similarity 0.86)
- Fast Image Processing With Fully-Convolutional Networks (ICCV 2017 · similarity 0.85)
- Accelerating CNN Training by Pruning Activation Gradients (ECCV 2020 · similarity 0.85)
- Accelerating Augmentation Invariance Pretraining (NeurIPS 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.
