This repository contains an exploratory implementation of a PointNet-style variational autoencoder for generating 3D chair geometry represented as point clouds.
The experimental pipeline includes:
- normalization of ModelNet chair meshes;
- surface sampling into point clouds;
- PointNet-style encoding;
- variational latent-space learning;
- generation of new point clouds;
- preliminary evaluation using Chamfer Distance;
- Poisson surface reconstruction.
Project status: experimental prototype. The model captures coarse chair-like structure but poorly preserves thin components such as legs. Surface reconstruction produces substantial artifacts and is not suitable for engineering use.
Random samples decoded from the latent space. The generated point clouds capture coarse seat-and-back structure, but point density is uneven and thin elements are poorly represented.
- Dataset: ModelNet, chair category
- Training objects: approximately 890
- Stored points per object: 16,384
- Encoder input points: 4,096
- Latent dimension: 256
- Optimizer: Adam
- Learning rate:
1e-4 - KL coefficient:
1e-4 - Training duration: 25 epochs
- Selected checkpoint: epoch 23
Training continued for 25 epochs, but the checkpoint from epoch 23 was selected because validation Chamfer Distance deteriorated during the final two epochs.
Ten generated point clouds were compared with the held-out validation subset.
| Metric | Value |
|---|---|
| Mean nearest-reference Chamfer-L2 | 0.0375 |
| Minimum nearest-reference Chamfer-L2 | 0.0306 |
| Maximum nearest-reference Chamfer-L2 | 0.0431 |
| Mean pairwise Chamfer-L2 | 0.0306 |
The reported metric uses non-squared Euclidean distances. These measurements provide a preliminary estimate of geometric similarity but do not establish perceptual quality, novelty or correct structural composition.
Generated point clouds were converted into polygonal meshes using Poisson Surface Reconstruction.
Poisson reconstruction preserves some coarse seat and back regions but loses chair legs and introduces severe surface artifacts.
The reconstruction results were unsatisfactory because the generated point clouds have uneven density, especially around thin structural elements.
- small dataset relative to model capacity;
- heavily overparameterized fully connected decoder;
- uneven generated point density;
- limited evaluation protocol;
- no independent final test set;
- loss of thin components during surface reconstruction;
- full historical training logs and trained weights are not included.
- compare the VAE with point-cloud diffusion models;
- replace the fully connected decoder with a more efficient coarse-to-fine or folding-based architecture;
- add density-aware and part-aware objectives;
- evaluate fidelity and coverage on a separate test split;
- explore implicit-surface and mesh-native decoders.