Research map / 3D vision / Face and shape
Holistic Geometric Feature Learning for Structured Reconstruction
Ziqiong Lu, Linxi Huan, Qiyuan Ma, Xianwei Zheng
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 frequency-domain feature learning (F-Learn) after first conv of HEAT | HEAT (ResNet50) · SpaceNet roof graphs | region F1 (%): 70.0 -> 73.4 | Improved |
| replace frequency-domain module with space-domain BConv/cascade/parallel modules (ablation) | HEAT · SpaceNet roof graphs | edge F1 (%): F-Learn 79.1 vs best space-domain 77.1 | Improved |
| space-domain conv modules added in place of F-Learn (ablation) | HEAT · SpaceNet roof graphs | corner F1 (%): 87.1 -> 88.1 (Cas.), 87.4 (BConv), 87.7 (Par.) vs F-Learn 88.6 | Mixed |
| apply F-Learn at deeper, lower-resolution layer (L2, 32x32) instead of first conv (ablation) | HEAT · SpaceNet roof graphs | region F1 (%): 73.4 -> 69.0 (HEAT baseline 70.0) | Hurt |
| apply F-Learn at L1 (64x64) instead of L0 (ablation) | HEAT · SpaceNet roof graphs | region F1 (%): 73.4 -> 70.9 | Hurt |
| add F-Learn to HEAT with ResNet18 backbone (ablation) | HEAT ResNet18 · SpaceNet roof graphs | region F1 (%): 68.2 -> 69.3 | Improved |
| add F-Learn to HEAT with ResNet34 backbone (ablation) | HEAT ResNet34 · SpaceNet roof graphs | region F1 (%): 68.0 -> 69.6 | Improved |
| add F-Learn with ground-truth corners (topology inference only) (ablation) | HEAT · SpaceNet roof graphs (GT corners) | edge F1 (%): 88.5 -> 91.1 | Improved |
| add space-domain BConv instead of F-Learn with GT corners (ablation) | HEAT · SpaceNet roof graphs (GT corners) | region F1 (%): 81.2 -> 79.6 | Hurt |
| add F-Learn module | HEAT · SpaceNet roof graphs | parameters and time per image: 68.747M -> 68.847M params; 0.10s -> 0.11s | No clear effect |
Closest papers by meaning
- Bridging the Geometry Mismatch: Frequency-Aware Anisotropic Serialization for Thin-Structure SSMs (ECCV 2026 · similarity 0.86)
- LIST: Learning Implicitly from Spatial Transformers for Single-View 3D Reconstruction (ICCV 2023 · similarity 0.86)
- D2IM-Net: Learning Detail Disentangled Implicit Fields From Single Images (CVPR 2021 · similarity 0.86)
- Learning Probabilistic Topological Representations Using Discrete Morse Theory (ICLR 2023 · similarity 0.85)
- Sharpening Neural Implicit Functions with Frequency Consolidation Priors (AAAI 2025 · similarity 0.84)
- DeepBranchTracer: A Generally-Applicable Approach to Curvilinear Structure Reconstruction Using Multi-Feature Learning (AAAI 2024 · 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.
