Research map / Point clouds / Point cloud classification
Variance-Aware Weight Initialization for Point Convolutional Neural Networks
Pedro Hermosilla, Michael Schelling, Tobias Ritschel, Timo Ropinski
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 |
|---|---|---|---|
| use variance-aware init instead of BatchNorm at batch size 2 | PCCNN · ScanObjectNN (batch size 2) | accuracy: 23.3 -> 85.7 | Improved |
| use variance-aware init instead of BatchNorm at batch size 16 | PCCNN · ScanObjectNN (batch size 16) | accuracy: 87.9 -> 83.5 | Hurt |
| use variance-aware init instead of BatchNorm at batch size 2 | MCConv · ScanObjectNN (batch size 2) | accuracy: 21.3 -> 86.3 | Improved |
| add variance-aware init (vs He init, no BN) (ablation) | KPConv · ScanObjectNN (batch size 2) | accuracy: 78.9 -> 84.8 | Improved |
| add variance-aware init (vs He init, no BN) (ablation) | PCCNN · ScanObjectNN (batch size 2) | accuracy: 13.5 -> 85.7 | Improved |
| use variance-aware init instead of Group Normalization | SPHConv · ScanObjectNN (batch size 2) | accuracy: 79.8 -> 82.2 | Improved |
| use variance-aware init (direct) instead of BatchNorm at batch size 1 | PointConv · ScanNet val (batch size 1) | mIoU: 41.4 -> 62.3 | Improved |
| add variance-aware init (vs standard init, no BN) (ablation) | PointConv · ScanNet val (batch size 1) | mIoU: 1.6 -> 62.3 | Improved |
| use transfer init instead of direct init (ablation) | MCConv · ScanNet val (batch size 1) | mIoU: 67.4 -> 67.9 | No clear effect |
| use variance-aware init instead of BatchNorm with batch 4 small clouds | SPHConv · ScanNet val | mIoU: 59.5 -> 60.1 | No clear effect |
| use gradient accumulation (virtual batch 4) with BN | KPConv · ScanNet val | mIoU: 32.1 | Hurt |
| add variance-aware init (vs no init, no BN) on official architecture (ablation) | PointConv · ScanNet val (official arch) | mIoU: 1.7 -> 59.3 | Improved |
Closest papers by meaning
- Norm matters: efficient and accurate normalization schemes in deep networks (NeurIPS 2018 · similarity 0.88)
- Four Things Everyone Should Know to Improve Batch Normalization (ICLR 2020 · similarity 0.88)
- Weight Normalization: A Simple Reparameterization to Accelerate Training of Deep Neural Networks (NeurIPS 2016 · similarity 0.88)
- Demystifying Batch Normalization in ReLU Networks: Equivalent Convex Optimization Models and Implicit Regularization (ICLR 2022 · similarity 0.88)
- Delving into Variance Transmission and Normalization: Shift of Average Gradient Makes the Network Collapse (AAAI 2021 · similarity 0.87)
- Revisiting Batch Norm Initialization (ECCV 2022 · similarity 0.87)
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.
