Research map / Detection and segmentation / CNN features
ThunderNet: Towards Real-Time Generic Object Detection on Mobile Devices
Zheng Qin, Zeming Li, Zhaoning Zhang, Yiping Bao, Gang Yu, Yuxing Peng, Jian 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 Context Enhancement Module (ablation) | compressed Light-Head R-CNN with SNet146 · COCO test-dev | AP: 21.6 -> 23.3 | Improved |
| add Spatial Attention Module (without CEM) (ablation) | compressed Light-Head R-CNN with SNet146 · COCO test-dev | AP: 21.6 -> 23.0 | Improved |
| add both CEM and SAM (ablation) | ThunderNet SNet146 · COCO test-dev | AP: 21.9 -> 23.7 | Improved |
| replace 3x3 conv in RPN with 5x5 depthwise plus 1x1 conv (ablation) | Light-Head R-CNN SNet146 · COCO test-dev | AP: 21.9 -> 21.8 | No clear effect |
| halve fc layer in R-CNN subnet to 1024-d (ablation) | Light-Head R-CNN SNet146 · COCO test-dev | AP: 21.8 -> 21.6 | No clear effect |
| use 5x5 instead of 3x3 depthwise convolutions (ablation) | SNet146 · COCO test-dev | AP: 22.7 -> 23.7 | Improved |
| use 5x5 depthwise instead of double 3x3 depthwise (ablation) | SNet146 · COCO test-dev | AP: 23.3 -> 23.7 | Improved |
| add 1024-d Conv5 to SNet146 (ablation) | SNet146 · COCO test-dev | AP: 23.7 -> 23.2 | Hurt |
| remove Conv5 from SNet49 (ablation) | SNet49 · COCO test-dev | AP: 19.2 -> 18.2 | Hurt |
| widen Conv5 of SNet49 to 1024-d (ablation) | SNet49 · COCO test-dev | AP: 19.2 -> 18.8 | Hurt |
| match input resolution to backbone size (SNet146 at 320 instead of 224) (ablation) | ThunderNet · COCO test-dev | AP: 18.7 -> 23.7 | Improved |
| use large-backbone-small-head instead of small-backbone-large-head (ablation) | ThunderNet · COCO test-dev | AP: 20.2 -> 23.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Pelee — Pelee: A Real-Time Object Detection System on Mobile Devices (NeurIPS 2018) · better on 2 of 2 shared results
- R-FCN — R-FCN: Object Detection via Region-based Fully Convolutional Networks (NeurIPS 2016) · better on 1 of 1 shared result
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
- MobileDets: Searching for Object Detection Architectures for Mobile Accelerators (CVPR 2021 · similarity 0.87)
- HBONet: Harmonious Bottleneck on Two Orthogonal Dimensions (ICCV 2019 · similarity 0.86)
- MnasFPN: Learning Latency-Aware Pyramid Architecture for Object Detection on Mobile Devices (CVPR 2020 · similarity 0.86)
- SmartAdapt: Multi-Branch Object Detection Framework for Videos on Mobiles (CVPR 2022 · similarity 0.86)
- Pelee: A Real-Time Object Detection System on Mobile Devices (NeurIPS 2018 · similarity 0.86)
- Efficient Event-Based Object Detection: A Hybrid Neural Network with Spatial and Temporal Attention (CVPR 2025 · 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.
