Research map / Architectures and efficiency / CNN design
GhostNet: More Features From Cheap Operations
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace VGG-16 convolutional layers with Ghost modules at s=2 | VGG-16 · CIFAR-10 | accuracy: 93.6 -> 93.7 | Helped |
| Replace ResNet-56 convolutional layers with Ghost modules at s=2 | ResNet-56 · CIFAR-10 | accuracy: 93.0 -> 92.7 | Hurt |
| Replace ResNet-50 convolutional layers with Ghost modules at s=2 | ResNet-50 · ImageNet ILSVRC-2012 | top-1 accuracy: 75.3 -> 75.0 | Hurt |
| Replace ResNet-50 convolutional layers with Ghost modules at s=4 | ResNet-50 · ImageNet ILSVRC-2012 | top-1 accuracy: 75.3 -> 74.1 | Hurt |
| Use GhostNet instead of MobileNetV3 at the listed comparable model sizes | ImageNet classification · ImageNet ILSVRC-2012 | top-1 accuracy: 65.4 -> 66.2; 73.3 -> 73.9; 75.2 -> 75.7 | Helped |
| Use GhostNet 1.1× as the detector backbone instead of MobileNetV2 1.0× | RetinaNet · MS COCO minival | mAP: 26.7% -> 26.6% | Mixed |
| Use GhostNet 1.1× as the detector backbone instead of MobileNetV2 1.0× | Faster R-CNN · MS COCO minival | mAP: 27.5% -> 26.9% | Hurt |
| Use transformation kernel d=1 instead of d=3 (ablation) | Ghost-VGG-16 · CIFAR-10 | accuracy: 93.7 -> 93.5 | Hurt |
| Use transformation kernel d=5 instead of d=3 (ablation) | Ghost-VGG-16 · CIFAR-10 | accuracy: 93.7 -> 93.4 | Hurt |
| Use transformation kernel d=7 instead of d=3 (ablation) | Ghost-VGG-16 · CIFAR-10 | accuracy: 93.7 -> 93.1 | Hurt |
| Increase Ghost module feature-generation ratio from s=2 to s=3 (ablation) | Ghost-VGG-16 · CIFAR-10 | accuracy: 93.7 -> 93.4 | Hurt |
| Increase Ghost module feature-generation ratio from s=2 to s=4 (ablation) | Ghost-VGG-16 · CIFAR-10 | accuracy: 93.7 -> 93.0 | Hurt |
Closest papers by meaning
- GhostNetV2: Enhance Cheap Operation with Long-Range Attention (NeurIPS 2022 · similarity 0.89)
- Scalable and Efficient Training of Large Convolutional Neural Networks with Differential Privacy (NeurIPS 2022 · similarity 0.84)
- Learning Transferable Adversarial Examples via Ghost Networks (AAAI 2020 · similarity 0.84)
- Beyond Filters: Compact Feature Map for Portable Deep Model (ICML 2017 · similarity 0.83)
- Rethinking Depthwise Separable Convolutions: How Intra-Kernel Correlations Lead to Improved MobileNets (CVPR 2020 · similarity 0.82)
- LSNet: See Large, Focus Small (CVPR 2025 · similarity 0.82)
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.