Research map / Detection and segmentation / CNN features
DiCENet: Dimension-Wise Convolutions for Efficient Networks
Sachin Mehta, Hannaneh Hajishirzi, Mohammad Rastegari
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 DimConv in place of depth-wise convolution (ablation) | ShuffleNetv2 with point-wise fusion · ImageNet val (120-150 MFLOPs) | top-1 accuracy: 68.14 -> 68.45 (146 M -> 135 M FLOPs) | Improved |
| add DimConv in place of depth-wise convolution (ablation) | ResNet with point-wise fusion · ImageNet val (120-150 MFLOPs) | top-1 accuracy: 67.80 -> 68.10 (142 M -> 132 M FLOPs) | Improved |
| replace point-wise convolution with DimFuse after DimConv (ablation) | ShuffleNetv2 · ImageNet val (120-150 MFLOPs) | top-1 accuracy / FLOPs: 68.45 -> 68.21 (135 M -> 122 M FLOPs) | Mixed |
| replace DimConv with depth-wise convolution when using DimFuse (ablation) | ShuffleNetv2 · ImageNet val (120-150 MFLOPs) | top-1 accuracy: 66.80 -> 68.21 | Improved |
| replace point-wise convolution with SE unit as fusion (ablation) | ShuffleNetv2 with DimConv · ImageNet val (120-150 MFLOPs) | top-1 accuracy: 68.45 -> 65.40 | Hurt |
| add SE unit on top of depth-wise plus point-wise convolution (ablation) | ResNet · ImageNet val (120-150 MFLOPs) | top-1 accuracy: 67.80 -> 68.20 | Improved |
| replace first point-wise convolution with DimFuse at same width (ablation) | ShuffleNetv2 DiCE unit · ImageNet val (120-150 MFLOPs) | top-1 accuracy: 68.21 -> 63.70 (122 M -> 72 M FLOPs) | Hurt |
| widen network 4x after replacing first point-wise with DimFuse (ablation) | ShuffleNetv2 DiCE unit · ImageNet val (120-150 MFLOPs) | top-1 accuracy: 63.70 -> 69.20 | Improved |
| replace separable convolution with the DiCE unit | MobileNet · ImageNet val (25-60 MFLOPs) | top-1 accuracy: 49.80 -> 52.55 | Improved |
| replace separable convolution with the DiCE unit | ShuffleNetv2 · ImageNet val (25-60 MFLOPs) | top-1 accuracy: 59.69 -> 62.80 | Improved |
| replace separable convolution with the DiCE unit | ResNet · ImageNet val (120-170 MFLOPs) | top-1 accuracy: 67.80 -> 67.90 (142 M -> 122 M FLOPs) | No clear effect |
| replace separable convolution with the DiCE unit | ShuffleNetv2 · ImageNet val (120-170 MFLOPs) | top-1 accuracy: 68.14 -> 68.21 (146 M -> 122 M FLOPs) | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ESPNetv2 — ESPNet: Efficient Spatial Pyramid of Dilated Convolutions for Semantic Segmentation (ECCV 2018) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DICE: Diversity in Deep Ensembles via Conditional Redundancy Adversarial Estimation (ICLR 2021 · similarity 0.85)
- EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks (ICML 2019 · similarity 0.84)
- ChannelNets: Compact and Efficient Convolutional Neural Networks via Channel-Wise Convolutions (NeurIPS 2018 · similarity 0.84)
- Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNet (ICLR 2019 · similarity 0.84)
- Bilinear Convolutional Neural Networks for Fine-Grained Visual Recognition (TPAMI 2017 · similarity 0.83)
- MogaNet: Multi-order Gated Aggregation Network (ICLR 2024 · similarity 0.83)
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.
