Research map / Detection and segmentation / CNN features
Pelee: A Real-Time Object Detection System on Mobile Devices
Robert J. Wang, Xiang Li, Charles X. Ling
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 transition layers without compression (ablation) | DenseNet-41 · Stanford Dogs (customized ILSVRC subset) | Top-1 accuracy: 75.02 -> 76.1 | Improved |
| Add post-activation to the no-compression transition variant (ablation) | DenseNet-41 · Stanford Dogs (customized ILSVRC subset) | Top-1 accuracy: 76.1 -> 75.2 | Hurt |
| Add dynamic bottleneck channels to the no-compression transition variant (ablation) | DenseNet-41 · Stanford Dogs (customized ILSVRC subset) | Top-1 accuracy: 75.02 -> 75.8 | Improved |
| Add a stem block to the no-compression transition and dynamic bottleneck variant (ablation) | DenseNet-41 · Stanford Dogs (customized ILSVRC subset) | Top-1 accuracy: 75.8 -> 76.8 | Improved |
| Add two-way dense layers (ablation) | PeleeNet variant without the extra dense layers · Stanford Dogs (customized ILSVRC subset) | Top-1 accuracy: 76.8 -> 78.8 | Improved |
| Add three extra dense layers (ablation) | PeleeNet · Stanford Dogs (customized ILSVRC subset) | Top-1 accuracy: 78.8 -> 79.25 | Improved |
| Add residual prediction blocks (ablation) | Pelee object detector · PASCAL VOC 2007 | mAP: 68.6 -> 70.8 | Improved |
| Use 1x1 instead of 3x3 prediction kernels (ablation) | Pelee object detector with residual prediction blocks · PASCAL VOC 2007 | mAP: 70.8 -> 70.9 | No clear effect |
| Use 1x1 instead of 3x3 prediction kernels (ablation) | Pelee object detector with residual prediction blocks · PASCAL VOC 2007 | FLOPs: 1,470 M -> 1,210 M | Improved |
| Omit the 38x38 feature map (ablation) | Pelee object detector without residual prediction blocks · PASCAL VOC 2007 | mAP: 69.3 -> 68.6 | Hurt |
| Use conventional convolution-based PeleeNet rather than MobileNet | ImageNet classification · ImageNet ILSVRC 2012 | Top-1 accuracy: 70.6 -> 72.6 | Improved |
| Use PeleeNet rather than MobileNet on TX2 at 224x224 | ImageNet classification · TX2, 224x224 | Speed (images per second): 136.2 -> 240.3 | Improved |
Later papers that compared with this method
1 later accepted paper reports Pelee as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- ThunderNet: Towards Real-Time Generic Object Detection on Mobile Devices (ICCV 2019 · better on 2 of 2 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
- MobileDets: Searching for Object Detection Architectures for Mobile Accelerators (CVPR 2021 · similarity 0.87)
- ThunderNet: Towards Real-Time Generic Object Detection on Mobile Devices (ICCV 2019 · similarity 0.86)
- YOLObile: Real-Time Object Detection on Mobile Devices via Compression-Compilation Co-Design (AAAI 2021 · similarity 0.85)
- ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices (CVPR 2018 · similarity 0.85)
- Mobile Video Object Detection With Temporally-Aware Feature Maps (CVPR 2018 · similarity 0.85)
- Speed/Accuracy Trade-Offs for Modern Convolutional Object Detectors (CVPR 2017 · 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.
