Research map / Detection and segmentation / CNN features
Beyond Fixed Grid: Learning Geometric Image Representation with a Deformable Grid
Jun Gao, Zian Wang, Jinchen Xuan, Sanja Fidler
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 |
|---|---|---|---|
| replace fixed max pooling with grid max pooling on the deformed grid | Modified ResNet50 · Cityscapes semantic segmentation (1/8 downsampling) | mIoU: 59.83 -> 64.75 | Improved |
| replace fixed average pooling with grid average pooling | Modified ResNet50 · Cityscapes semantic segmentation (1/32 downsampling) | mIoU: 43.54 -> 48.30 | Improved |
| replace fixed pooling with grid pooling, measured on boundary quality | Modified ResNet50 · Cityscapes semantic segmentation (1/16 downsampling) | boundary F (4px): 49.00 -> 53.98 | Improved |
| add learned grid deformation instead of keeping the grid fixed (ablation) | DefGrid annotation model · Cityscapes-Multicomp | F2: 74.77 -> 77.04 | Improved |
| add learned grid deformation instead of keeping the grid fixed (ablation) | DefGrid annotation model · Cityscapes-Multicomp | mIoU: 81.67 -> 82.17 | Improved |
| reason on the deformed grid instead of Curve-GCN polygon | Curve-GCN · ssTEM (cross-domain) | F2: 59.85 -> 65.14 | Improved |
| reason on the deformed grid instead of Curve-GCN polygon, interactive annotation | Curve-GCN · Cityscapes interactive annotation (round 0) | F1: 58.01 -> 61.43 | Improved |
| predict per-cell labels on the deformed grid instead of pixel-wise labels | pixel-wise annotation model · Cityscapes-Stretch | mIoU: 80.01 -> 81.42 | Improved |
| use L1 instead of L2 distance in reconstruction loss (ablation) | DefGrid · training loss | quality | Improved |
| add DefGrid for unsupervised superpixels versus SEAL supervised | superpixel baselines · BSDS500 test | boundary precision | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PolyTransform — PolyTransform: Deep Polygon Transformer for Instance Segmentation (CVPR 2020) · better on 1 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
- Multigrid Neural Architectures (CVPR 2017 · similarity 0.87)
- Deformable Convolutional Networks (ICCV 2017 · similarity 0.87)
- Deformable Kernels: Adapting Effective Receptive Fields for Object Deformation (ICLR 2020 · similarity 0.87)
- PointGrid: A Deep Network for 3D Shape Understanding (CVPR 2018 · similarity 0.87)
- Cascade Image Matting With Deformable Graph Refinement (ICCV 2021 · similarity 0.86)
- DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs (TPAMI 2017 · 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.
