Research map / Detection and segmentation / CNN features
ScratchDet: Training Single-Shot Object Detectors From Scratch
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 |
|---|---|---|---|
| add BatchNorm throughout the backbone and detection head (ablation) | SSD trained from scratch · PASCAL VOC 2007 test | mAP: 67.6 -> 78.7 | Helped |
| add BatchNorm in the backbone and use learning rate 0.05 (ablation) | SSD trained from scratch · PASCAL VOC 2007 test | mAP: 67.6% to 78.0% | Helped |
| add BatchNorm in the detection head (ablation) | SSD trained from scratch · PASCAL VOC 2007 test | mAP: 67.6% to 71.0% | Helped |
| increase learning rate from 0.001 to 0.01 with head BatchNorm (ablation) | SSD trained from scratch · PASCAL VOC 2007 test | mAP: 71.0% to 75.6% | Helped |
| increase learning rate from 0.01 to 0.05 with whole-network BatchNorm (ablation) | SSD trained from scratch · PASCAL VOC 2007 test | mAP: 77.3% to 78.7% | Helped |
| add BatchNorm to SSD initialized from pretrained VGG-16 (ablation) | pretrained SSD · PASCAL VOC 2007 test | mAP: 77.2% v.s 78.2% | Helped |
| remove all BatchNorm layers (ablation) | DSOD · PASCAL VOC 2007 test | mAP: 77.7% to 71.5% | Hurt |
| change the first convolution kernel from 7×7 to 3×3 while retaining downsampling (ablation) | ResNet-18 SSD · PASCAL VOC 2007 test | mAP: 73.1% for 7×7 v.s 73.2% for 3×3 | No clear effect |
| remove downsampling in the first convolution layer (ablation) | ResNet-18 SSD · PASCAL VOC 2007 test | mAP: 73.1% to 77.6% | Helped |
| remove only the second downsampling operation, retaining first-layer stride 2 (ablation) | ResNet-18 SSD · PASCAL VOC 2007 test | mAP: 73.1% to 75.3% | Helped |
| increase root-block depth from one to three 3×3 convolutions (ablation) | ResNet-18-B SSD · PASCAL VOC 2007 test | mAP: 77.8% to 78.5% | Helped |
| increase root-block depth beyond three convolutions (ablation) | ResNet-18-B SSD · PASCAL VOC 2007 test | mAP: 78.5% (3 layers); 78.4% (4 layers); 78.5% (5 layers) | Mixed |
Closest papers by meaning
- Training Object Detectors From Scratch: An Empirical Study in the Era of Vision Transformer (CVPR 2022 · similarity 0.88)
- MegDet: A Large Mini-Batch Object Detector (CVPR 2018 · similarity 0.88)
- M2Det: A Single-Shot Object Detector Based on Multi-Level Feature Pyramid Network (AAAI 2019 · similarity 0.87)
- Training-Time-Friendly Network for Real-Time Object Detection (AAAI 2020 · similarity 0.87)
- iDETR: Implicit DETR for Tiny Object Detection (NeurIPS 2026 · similarity 0.86)
- DetKDS: Knowledge Distillation Search for Object Detectors (ICML 2024 · similarity 0.86)
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.