Research map / Detection and segmentation / CNN features
Latency-aware Spatial-wise Dynamic Networks
Yizeng Han, Zhihang Yuan, Yifan Pu, Chenhao Xue, Shiji Song, Guangyu Sun, Gao Huang
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 |
|---|---|---|---|
| fuse masker and first conv (ablation) | LAS-ResNet-101 block · stage-1 block, V100 | latency (us): 163.2 -> 90.1 | Improved |
| fuse gather with dynamic conv (ablation) | LAS-ResNet-101 block · stage-1 block, V100 | latency (us): 90.1 -> 86.7 | Improved |
| fuse scatter with add (ablation) | LAS-ResNet-101 block · stage-1 block, V100 | latency (us): 86.7 -> 71.4 | Improved |
| use coarse spatial granularity S>1 instead of pixel-level S=1 (ablation) | LAS-ResNet-101 · ImageNet, V100 | latency | Improved |
| use coarse granularity (4-4-2-1, 8-4-7-1) vs 1-1-1-1 (ablation) | LAS-ResNet · ImageNet | accuracy at similar FLOPs | Improved |
| use LAS-ResNet-101 instead of ResNet-101 at t=0.4 | ResNet-101 · ImageNet | latency: -36% on V100, -46% on TX2 | Improved |
| replace S=4-4-2-1 with layer skipping S=56-28-14-7 (ablation) | LAS-ResNet-50 · ImageNet | accuracy (%): 76.6 -> 76.1 | Hurt |
| use LAS-ResNet-101 (t=0.6) as Faster R-CNN backbone | ResNet-101 · COCO Faster R-CNN | mAP (%): 39.4 -> 40.3 | Improved |
| use LAS-ResNet-101 (t=0.4) as Faster R-CNN backbone | ResNet-101 · COCO Faster R-CNN | backbone latency V100 (ms): 39.5 -> 25.3 | Improved |
| use LAS-ResNet-101 (t=0.4) as RetinaNet backbone | ResNet-101 · COCO RetinaNet | mAP (%): 38.5 -> 38.9 | No clear effect |
| use LAS-ResNet-101 (t=0.4, 4-4-2-1) in Mask R-CNN | ResNet-101 · COCO Mask R-CNN | AP mask (%): 36.1 -> 36.1 | No clear effect |
Closest papers by meaning
- Latency-Aware Unified Dynamic Networks for Efficient Image Recognition (TPAMI 2024 · similarity 0.91)
- Glance and Focus: a Dynamic Approach to Reducing Spatial Redundancy in Image Classification (NeurIPS 2020 · similarity 0.87)
- Exploring Hidden Dimensions in Accelerating Convolutional Neural Networks (ICML 2018 · similarity 0.86)
- HarDNet: A Low Memory Traffic Network (ICCV 2019 · similarity 0.86)
- Dynamic Neural Networks: A Survey (TPAMI 2021 · similarity 0.86)
- Efficient Latency-Aware CNN Depth Compression via Two-Stage Dynamic Programming (ICML 2023 · similarity 0.85)
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.
