Research map / Detection and segmentation / CNN features
Accelerating Very Deep Convolutional Networks for Classification and Detection
Xiangyu Zhang, Jianhua Zou, Kaiming He, Jian Zhong Sun
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 |
|---|---|---|---|
| add asymmetric reconstruction (account for approximated previous layers) (ablation) | low-rank layer approximation · SPP-10, 3 layers approximated at 4x | top-5 error increase: over 1.0% | Improved |
| add 3d decomposition (spatial plus channel) to asymmetric reconstruction (ablation) | low-rank acceleration · SPP-10 ImageNet 4x | top-5 error (1-view): 16.7 -> 14.1 | Improved |
| add fine-tuning to asym. (3d) | low-rank acceleration · SPP-10 ImageNet 4x | top-5 error (1-view): 14.1 -> 13.8 | Improved |
| add whole-model rank selection (ablation) | asymmetric solver · SPP-10 ImageNet 4x | increase of top-5 error: 4.32 -> 4.20 | Improved |
| add whole-model rank selection (ablation) | asymmetric solver · VGG-16 ImageNet 4x | increase of top-5 error: 6.38 -> 3.84 | Improved |
| add whole-model rank selection (ablation) | asymmetric solver · VGG-16 ImageNet 2x | increase of top-5 error: 0.99 -> 0.28 | Improved |
| use nonlinear solution instead of linear solution (ablation) | single-layer low-rank approximation · SPP-10 single-layer | increase of top-5 error | Improved |
| use our method instead of Jaderberg et al. spatial decomposition | whole-model acceleration · SPP-10 ImageNet 4x | increase of top-5 error: 6.0% -> 4.2% | Improved |
| use our asym. (3d) instead of Jaderberg et al. | VGG-16 acceleration · VGG-16 ImageNet 5x | increase of top-5 error: 29.7 -> 2.0 | Improved |
| use accelerated 3d model instead of same architecture trained from scratch (ablation) | decomposed SPP-10 architecture · SPP-10 ImageNet 4x | top-5 error (1-view): 16.9 -> 14.1 | Improved |
| increase conv speedup from 3x to 5x | accelerated VGG-16 + Fast R-CNN · PASCAL VOC 2007 | mAP: 66.9 -> 65.2 | Hurt |
| add fine-tuning to asym. (3d) | VGG-16 acceleration at 4x · ImageNet val | increase of top-5 error: 0.9 -> 0.3 | Improved |
Closest papers by meaning
- Fast Image Processing With Fully-Convolutional Networks (ICCV 2017 · similarity 0.88)
- DeepOBS: A Deep Learning Optimizer Benchmark Suite (ICLR 2019 · similarity 0.88)
- More ConvNets in the 2020s: Scaling up Kernels Beyond 51x51 using Sparsity (ICLR 2023 · similarity 0.88)
- SCNN: A General Distribution Based Statistical Convolutional Neural Network with Application to Video Object Detection (AAAI 2019 · similarity 0.88)
- Recurrent Residual Module for Fast Inference in Videos (CVPR 2018 · similarity 0.88)
- Accelerating Sparse Convolution with Column Vector-Wise Sparsity (NeurIPS 2022 · 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.
