Research map / Optimization / Implicit bias and generalization
Building Deep Networks on Grassmann Manifolds
Zhiwu Huang, Jiqing Wu, Luc Van Gool
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 a second Projection-Pooling block (ablation) | GrNet · AFEW, HDM05, PaSC control, and PaSC handheld | accuracy / verification performance: 32.08% -> 34.23%; 57.73%$\pm$2.24 -> 59.23%$\pm$1.78; 80.15% -> 80.52%; 72.51% -> 72.76% | Mixed |
| add Projection-Pooling blocks rather than use no blocks (ablation) | GrNet · AFEW, HDM05, PaSC control, and PaSC handheld | accuracy / verification performance: 25.34% -> 34.23%; 51.12%$\pm$3.55 -> 59.23%$\pm$1.78; 68.52% -> 80.52%; 63.92% -> 72.76% | Improved |
| remove ProjMap (Riemannian computing) (ablation) | GrNet · AFEW | accuracy: 17.62% | Hurt |
| remove ReOrth (geometry-aware learning) (ablation) | GrNet · AFEW | accuracy: 26.15% | Hurt |
| use multiple FRMap projections instead of a single projection (ablation) | GrNet · three visual recognition datasets | performance | Improved |
| use W-ProjPooling instead of A-ProjPooling (ablation) | GrNet with M-FRMap · three visual recognition datasets | performance | Improved |
| update FRMap weights on Stiefel manifolds instead of PSD manifolds (ablation) | GrNet · four reported dataset evaluations | performance: 32.13%, 57.25%, 80.15%, 71.28% | Hurt |
| use GrNet rather than the strongest reported baseline | GrNet-2Blocks · AFEW validation | accuracy: 34.23% -> 34.23% | No clear effect |
| use GrNet rather than the strongest reported baseline | GrNet-2Blocks · HDM05 | accuracy: 61.45%$\pm$1.12 -> 59.23%$\pm$1.78 | Hurt |
| use GrNet rather than the strongest reported baseline | GrNet-2Blocks · PaSC control | verification performance: 80.12% -> 80.52% | Improved |
| use GrNet rather than the strongest reported baseline | GrNet-2Blocks · PaSC handheld | verification performance: 72.83% -> 72.76% | No clear effect |
Closest papers by meaning
- Get the Best of Both Worlds: Improving Accuracy and Transferability by Grassmann Class Representation (ICCV 2023 · similarity 0.89)
- Grassmann Pooling as Compact Homogeneous Bilinear Pooling for Fine-Grained Visual Classification (ECCV 2018 · similarity 0.88)
- Matrix Manifold Neural Networks++ (ICLR 2024 · similarity 0.87)
- Learning Topology-Driven Multi-Subspace Fusion for Grassmannian Deep Networks (AAAI 2026 · similarity 0.87)
- Random features for Grassmannian kernel approximation with bounded rank-one projections (NeurIPS 2026 · similarity 0.86)
- Building Transformation Layers for Riemannian Neural Networks (NeurIPS 2026 · similarity 0.84)
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.
