How njuvision/sparsepcgc works
SparsePCGC is a learned point-cloud geometry compression system. Its user-facing workflow prepares voxelized sparse point clouds, estimates occupancy through a multiscale neural model, and supports dense and sparse LiDAR evaluation in lossless and lossy modes; offset prediction and G-PCC comparisons extend the workflow. Training is a separate path that feeds point-cloud data to the model and records checkpoints and metrics. The README documents the test modes and capabilities; source-index references support the internal links shown, while links from test entry points to model components are left out because those files were not sampled.
Main entry points: model.py Multiscale PCC model · data_loader.py Sparse data loader · autoencoder.py Sparse autoencoder · train.py Training entry point
Components 20
Point-cloud preparation
- Sparse data loader
data_utils/data_loader.py - Point-cloud datasets
data_utils/dataset.py - Point-cloud I/O
data_utils/inout.py - Voxel quantization
data_utils/quantize.py - Spatial partitioning
data_utils/partition.py - Sparse tensor ops
data_utils/sparse_tensor.py
Learned compression
- Multiscale PCC model
models/model.py - Sparse autoencoder
models/autoencoder.py - Occupancy decoder
models/decoder.py - Entropy model
models/entropy_model.py - Sparse residual blocks
models/resnet.py - Offset prediction
models/model_offset.py
Training and inference
- Test entry points
test - Training entry point
train/train.py - Training lifecycle
train/trainer_base.py
Codec evaluation
- Point-cloud metrics
extension/metrics.py - G-PCC and distortion tools
extension/tools.py - G-PCC wrapper
extension/gpcc.py
Other components
- Researcher
- Point-cloud files
Connections 33
- Researcher → Test entry points: runs tests
README.md - Researcher → Training entry point: starts training
README.md - Point-cloud files → Sparse data loader: loads coordinates
data_utils/data_loader.py - Point-cloud files → Point-cloud datasets: supplies samples
- Sparse data loader → Point-cloud I/O: reads files
data_utils/data_loader.py - Sparse data loader → Voxel quantization: quantizes points
data_utils/data_loader.py - Sparse data loader → Spatial partitioning: partitions points
data_utils/data_loader.py - Point-cloud datasets → Point-cloud I/O: reads files
data_utils/dataset.py - Point-cloud datasets → Voxel quantization: quantizes points
data_utils/dataset.py - Point-cloud datasets → Spatial partitioning: partitions points
data_utils/dataset.py - Multiscale PCC model → Sparse data loader: loads sparse input
models/model.py - Multiscale PCC model → Voxel quantization: quantizes scales
models/model.py - Multiscale PCC model → Point-cloud I/O: writes output
models/model.py - Multiscale PCC model → Point-cloud metrics: measures distortion
models/model.py - Multiscale PCC model → Sparse autoencoder: selects encoder
models/model.py - Multiscale PCC model → Occupancy decoder: selects decoder
models/model.py - Sparse autoencoder → Sparse residual blocks: uses blocks
models/autoencoder.py - Sparse autoencoder → Entropy model: uses coding model
models/autoencoder.py - Sparse autoencoder → Occupancy decoder: uses decoder
models/autoencoder.py - Sparse autoencoder → Sparse tensor ops: uses tensor ops
models/autoencoder.py - Occupancy decoder → Sparse residual blocks: uses blocks
models/decoder.py - Occupancy decoder → Sparse tensor ops: uses tensor ops
models/decoder.py - Training entry point → Sparse data loader: loads batches
train/train.py - Training entry point → Multiscale PCC model: trains model
train/train.py - Training entry point → Training lifecycle: extends trainer
train/train.py - Training entry point → Sparse tensor ops: builds sparse input
train/train.py - Offset prediction → Sparse residual blocks: uses blocks
models/model_offset.py - Offset prediction → G-PCC and distortion tools: uses codec tools
models/model_offset.py - Offset prediction → Voxel quantization: quantizes coordinates
models/model_offset.py - Offset prediction → Sparse data loader: loads input
models/model_offset.py - Test entry points → Multiscale PCC model: runs compression
- Test entry points → Offset prediction: runs offset mode
- G-PCC and distortion tools → G-PCC wrapper: invokes wrapper