Research map / Detection and segmentation / CNN features
DSOD: Learning Deeply Supervised Object Detectors From Scratch
Zhiqiang Shen, Zhuang Liu, Jianguo Li, Yu-Gang Jiang, Yurong Chen, Xiangyang Xue
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 w/o pooling layer (4 dense blocks) (ablation) | DSOD300 DS/32-12-16-0.5 · VOC 2007 test (07+12) | mAP: 59.9 -> 61.6 | Improved |
| use compression factor 1 instead of 0.5 (ablation) | DSOD300 backbone · VOC 2007 test (07+12) | mAP: 61.6 -> 64.5 | Improved |
| widen bottleneck layers (ablation) | DSOD300 backbone · VOC 2007 test (07+12) | mAP: 64.5 -> 68.6 | Improved |
| widen first conv layer (ablation) | DSOD300 backbone · VOC 2007 test (07+12) | mAP: 68.6 -> 69.7 | Improved |
| increase growth rate from 16 to 48 (ablation) | DSOD300 backbone · VOC 2007 test (07+12) | mAP: 69.7 -> 74.5 | Improved |
| add stem block (ablation) | DSOD300 DS/64-192-48-1 · VOC 2007 test (07+12) | mAP: 74.5 -> 77.3 | Improved |
| use dense prediction layers instead of plain (ablation) | DSOD300 DS/64-192-48-1 · VOC 2007 test (07+12) | mAP: 77.3 -> 77.7 | Improved |
| use dense prediction layers instead of plain (ablation) | DSOD300 DS/64-192-48-1 · VOC 2007 test (07+12) | speed (fps): 20.6 -> 17.4 | Hurt |
| use dense prediction layers instead of plain (ablation) | SSD300S VGGNet (from scratch) · VOC 2007 test (07+12) | mAP: 69.6 -> 70.4 | Improved |
| use dense prediction layers instead of plain (ablation) | SSD VGG-16 (pretrained) · VOC 2007 test (07+12) | mAP: 75.8 -> 76.1 | Improved |
| add deep supervision via dense layer-wise connections (DSOD vs SSD from scratch) (ablation) | SSD300S VGG16 from scratch · VOC 2007 test (07+12) | mAP: 69.6 -> 77.7 | Improved |
| train from scratch instead of fine-tuning ImageNet model (ablation) | SSD300 VGGNet · VOC 2007 test (07+12) | mAP: 75.8 -> 69.6 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- R-FCN — R-FCN: Object Detection via Region-based Fully Convolutional Networks (NeurIPS 2016) · better on 2 of 4 shared results
- Faster RCNN — Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (TPAMI 2016) · better on 3 of 3 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
- Object Detection from Scratch with Deep Supervision (TPAMI 2019 · similarity 0.98)
- Deep Learning for Generic Object Detection: A Survey (IJCV 2019 · similarity 0.90)
- Salient Object Detection in the Deep Learning Era: An In-Depth Survey (TPAMI 2021 · similarity 0.89)
- ScratchDet: Training Single-Shot Object Detectors From Scratch (CVPR 2019 · similarity 0.89)
- LEOD: Label-Efficient Object Detection for Event Cameras (CVPR 2024 · similarity 0.88)
- SimLTD: Simple Supervised and Semi-Supervised Long-Tailed Object Detection (CVPR 2025 · similarity 0.87)
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.
