How facebookresearch/deepsdf works
DeepSDF learns a continuous signed-distance representation of 3D shapes and turns it into reconstructed meshes. The main workflow preprocesses source meshes into SDF samples, trains a decoder and per-shape latent codes, reconstructs test shapes, then evaluates the resulting meshes against sampled ground-truth surfaces. The graph also includes the documented surface-sampling branch and experiment visualization. README statements establish the overall workflow; some internal links are dashed where the available excerpts do not show the call site.
Main entry points: train_deep_sdf.py Autodecoder training · reconstruct.py Latent optimization · evaluate.py Reconstruction evaluation · preprocess_data.py Preprocess orchestration
Components 21
Data preparation
- Preprocess orchestration
preprocess_data.py - SDF sampling
src/PreprocessMesh.cpp - Surface sampling
src/SampleVisibleMeshSurface.cpp - SDF and surface samples
- SDF dataset loader
deep_sdf/data.py
Model learning
- Experiment configuration
deep_sdf/workspace.py - Autodecoder training
train_deep_sdf.py - SDF decoder
networks/deep_sdf_decoder.py - Model and latent checkpoints
Reconstruction
- Latent optimization
reconstruct.py - Mesh extraction
deep_sdf/mesh.py - Reconstructed meshes
Evaluation and tools
- Reconstruction evaluation
evaluate.py - Chamfer metric
deep_sdf/metrics/chamfer.py - Evaluation results
- Training progress plot
plot_log.py - Training mesh generation
generate_training_meshes.py - Interior sample inspection
show_interior_samples.py - Experiment storage helpers
deep_sdf/workspace.py
Other components
- Researcher
- Source meshes
Connections 26
- Researcher → Preprocess orchestration: runs preprocessing
README.md - Source meshes → Preprocess orchestration: provides meshes
README.md - Preprocess orchestration → SDF sampling: invokes sampler
README.md - Preprocess orchestration → Surface sampling: requests surface samples
README.md - SDF sampling → SDF and surface samples: writes SDF samples
README.md - Surface sampling → SDF and surface samples: writes surface samples
README.md - SDF and surface samples → SDF dataset loader: provides training samples
- Researcher → Autodecoder training: starts training
README.md - Experiment configuration → Autodecoder training: configures training
README.md - SDF dataset loader → Autodecoder training: supplies batches
- Autodecoder training → SDF decoder: optimizes decoder
- Autodecoder training → Model and latent checkpoints: saves state
README.md - Researcher → Latent optimization: starts reconstruction
README.md - Model and latent checkpoints → Latent optimization: provides weights
README.md - SDF and surface samples → Latent optimization: provides test SDF
README.md - Latent optimization → SDF decoder: queries decoder
reconstruct.py - Latent optimization → Mesh extraction: extracts mesh
reconstruct.py - Mesh extraction → Reconstructed meshes: writes meshes
deep_sdf/mesh.py - Researcher → Reconstruction evaluation: starts evaluation
README.md - Reconstructed meshes → Reconstruction evaluation: provides reconstructions
README.md - SDF and surface samples → Reconstruction evaluation: provides ground truth
README.md - Reconstruction evaluation → Chamfer metric: computes distance
- Reconstruction evaluation → Evaluation results: writes metrics
README.md - Researcher → Training progress plot: views training progress
README.md - Researcher → Training mesh generation: generates training meshes
- Researcher → Interior sample inspection: inspects samples