Research map / Detection and segmentation / CNN features
Learning Affinity with Hyperbolic Representation for Spatial Propagation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 affinity conv layers with HAM (hyperbolic affinity module) | CSPN · NYUv2 depth completion | RMSE: 0.116 -> 0.102 | Helped |
| Replace affinity conv layers with HAM | CSPN · ScanNet depth completion | RMSE: 0.080 -> 0.073 | Helped |
| Replace affinity conv layers with HAM | CSPN · Virtual-KITTIv2 depth completion | RMSE: 12.233 -> 9.612 | Helped |
| Replace affinity conv layers with HAM | CSPN on KITTI-DC · KITTI depth completion | RMSE: 0.8229 -> 0.8182 | Helped |
| Replace affinity conv layers with HAM | DifNet-style SPN (DeepLabV2) · PASCAL VOC 2012 | mIoU: 70.02 -> 71.16 | Helped |
| Replace affinity with HAM | DifNet-style SPN · Pascal Context, NYUv2, ADE20K semantic segmentation | mIoU: Pascal Context 59.77->60.30; NYUv2 28.91->30.45; ADE20K 23.30->25.28 | Helped |
| Naive hyperbolic convolution (no beta-priority or geodesic weight) instead of HAM | CSPN depth completion and DifNet segmentation · NYUv2 depth and PASCAL VOC | RMSE / mIoU: Naive RMSE 0.108 vs CSPN 0.116 (better); Naive mIoU 70.49 vs DifNet 70.02 (better) but Naive pixel accuracy 91.53 vs bas | Mixed |
| Remove beta-priority from HAM (ablation) | HAM · PASCAL VOC 2012 (selected classes) | mIoU: 71.16 -> 70.77 when removed | Hurt |
| Remove geodesic weight from HAM (ablation) | HAM · PASCAL VOC 2012 | mIoU: 71.16 -> 70.91 when removed | Hurt |
| Remove beta-priority from HAM (ablation) | HAM · NYUv2 depth completion | RMSE: 0.1022 -> 0.1047 when removed | Hurt |
| Remove geodesic weight from HAM (ablation) | HAM · NYUv2 depth completion | RMSE: 0.1022 -> 0.1050 when removed | Hurt |
| Replace affinity layer with HAM at sparse seeds | NLSPN · NYUv2 depth completion, 100 samples | RMSE: 0.327 -> 0.258 | Helped |
Closest papers by meaning
- Hyperbolic Representation Learning: Revisiting and Advancing (ICML 2023 · similarity 0.89)
- Learning Affinity via Spatial Propagation Networks (NeurIPS 2017 · similarity 0.87)
- Fully Hyperbolic Convolutional Neural Networks for Computer Vision (ICLR 2024 · similarity 0.85)
- Accept the Modality Gap: An Exploration in the Hyperbolic Space (CVPR 2024 · similarity 0.85)
- Learning Along the Arrow of Time: Hyperbolic Geometry for Backward-Compatible Representation Learning (ICML 2025 · similarity 0.84)
- Learning Structured Representations with Hyperbolic Embeddings (NeurIPS 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.