Research map / Detection and segmentation / CNN features
SpineNet: Learning Scale-Permuted Backbone for Recognition and Localization
Xianzhi Du, Tsung-Yi Lin, Pengchong Jin, Golnaz Ghiasi, Mingxing Tan, Yin Cui, Quoc V. Le, Xiaodan Song
Check 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 |
|---|---|---|---|
| use a learned scale-permuted backbone with cross-scale connections instead of ResNet-FPN | RetinaNet · COCO val2017 | AP: 37.8 -> 40.7 | Improved |
| add scale and block-type adjustments to the scale-permuted architecture (ablation) | SpineNet-49 · COCO val2017 | AP: 40.7 -> 40.8 | Improved |
| replace learned feature-scale ordering with a fixed Hourglass ordering (ablation) | R0-SP53 · COCO val2017 | AP: 40.7 -> 38.3% | Hurt |
| replace learned feature-scale ordering with a fixed Fish ordering (ablation) | R0-SP53 · COCO val2017 | AP: 40.7 -> 37.5% | Hurt |
| remove short-range cross-scale connections (ablation) | R0-SP53 · COCO val2017 | AP: 40.7 -> 35.8% | Hurt |
| remove long-range cross-scale connections (ablation) | R0-SP53 · COCO val2017 | AP: 40.7 -> 28.6% | Hurt |
| replace cross-scale connections with sequential connections (ablation) | R0-SP53 · COCO val2017 | AP: 40.7 -> 28.2% | Hurt |
| allocate more feature layers to the scale-permuted network (ablation) | ResNet-to-SpineNet model family · COCO val2017 | AP: 37.8 -> 38.7 -> 39.7 -> 39.6 -> 40.7 | Improved |
| add stochastic depth, replace ReLU with swish, and use label smoothing of 0.1 (ablation) | SpineNet classification models · ImageNet ILSVRC-2012 and iNaturalist-2017 | Top-1 accuracy: 77.0 -> 78.1; 59.3 -> 63.3 | Improved |
| use SpineNet rather than ResNet for fine-grained classification | Image classification models · iNaturalist-2017 validation | Top-1 accuracy: 59.3 vs 54.6 | Improved |
| replace ResNet-FPN with SpineNet-49 | Mask R-CNN · COCO val2017, 640×640 | AP: 42.7 vs 42.9 | Improved |
| use a SpineNet MBConv backbone instead of mobile detector baselines | Mobile-size object detection · COCO test-dev | AP: 28.6 vs 25.7 | Improved |
Later papers that compared with this method
3 later accepted papers report SpineNet-96 as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- MAE-DET: Revisiting Maximum Entropy Principle in Zero-Shot NAS for Efficient Object Detection (ICML 2022 · better on 1 of 1 shared result)
- OPANAS: One-Shot Path Aggregation Network Architecture Search for Object Detection (CVPR 2021 · better on 1 of 1 shared result)
- Efficient Scale-Permuted Backbone with Learned Resource Distribution (ECCV 2020 · better on 0 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
- Efficient Scale-Permuted Backbone with Learned Resource Distribution (ECCV 2020 · similarity 0.91)
- DetNet: Design Backbone for Object Detection (ECCV 2018 · similarity 0.87)
- DetNAS: Backbone Search for Object Detection (NeurIPS 2019 · similarity 0.87)
- Res2Net: A New Multi-Scale Backbone Architecture (TPAMI 2019 · similarity 0.86)
- Scale-Aware Trident Networks for Object Detection (ICCV 2019 · similarity 0.86)
- DECO: Unleashing the Potential of ConvNets for Query-based Detection and Segmentation (ICLR 2025 · 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.
