Research map / 3D vision / Face and shape
Learning 3D Shapes as Multi-Layered Height-maps using 2D Convolutional Networks
Kripasindhu Sarkar, Basavaraj Hampiholi, Kiran Varanasi, Didier Stricker
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 element-wise max merge with concatenation followed by convolution (independent branches) (ablation) | 3-view MLH VGG16 · ModelNet40 | accuracy (%): 91.29 -> 93.11 | Improved |
| use independent (non-shared) view branches with max merge (ablation) | 3-view MLH VGG16 · ModelNet40 | accuracy (%): 91.25 -> 91.29 | No clear effect |
| initialize from ImageNet-pretrained VGG16 instead of random weights (ablation) | 3-view MLH VGG16 · ModelNet40 | accuracy (%): 89.63 -> 93.11 | Improved |
| copy pretrained RGB weights into 5-channel first layer instead of random first layer (ablation) | MLH VGG16 · ModelNet10/ModelNet40 | test accuracy: approximately 0.1% | Improved |
| increase number of MLH layers from 1 to 2 and up to 5 (ablation) | 3-view MLH merged model · ModelNet40 | accuracy: higher than 90.5% with 1 layer; big jump from layer 1 to 2, slow saturation till layer 5 | Improved |
| use concatenation+convolution merge with 1-layer descriptors instead of shared max (MVCNN style) (ablation) | 1-layer MLH, 3 views · ModelNet40 | accuracy (%): 89.78 -> 90.72 | Improved |
| use independent branches with max merge instead of shared with max for 1-layer descriptors (ablation) | 1-layer MLH, 3 views · ModelNet40 | accuracy (%): 89.78 -> 90.15 | Improved |
| replace VGG16 with ResNet50 or Inception models (ablation) | MLH classifier · ModelNet40 | test accuracy | No clear effect |
| use 3 MLH views with cat merge instead of 80-view MVCNN rendered images | shape classification · ModelNet40 | accuracy (%): 90.10 -> 93.11 | Improved |
| use single-view MLH 2D CNN instead of OctNet 256 voxel network | shape classification · ModelNet10 | accuracy (%): 90.3 -> 90.97 | Improved |
Closest papers by meaning
- MVTN: Learning Multi-view Transformations for 3D Understanding (IJCV 2024 · similarity 0.87)
- Neural Convolutional Surfaces (CVPR 2022 · similarity 0.87)
- Enhancing 2D Representation via Adjacent Views for 3D Shape Retrieval (ICCV 2019 · similarity 0.87)
- Learning View-Based Graph Convolutional Network for Multi-View 3D Shape Analysis (TPAMI 2022 · similarity 0.87)
- MLVCNN: Multi-Loop-View Convolutional Neural Network for 3D Shape Retrieval (AAAI 2019 · similarity 0.86)
- Synthesizing 3D Shapes via Modeling Multi-View Depth Maps and Silhouettes With Deep Generative Networks (CVPR 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.
