How yewzijian/regtr works
REGTR is a point-cloud registration system: it loads and transforms point-pair data, predicts point correspondences and confidence with a KPConv-plus-transformer network, and uses those predictions for registration and evaluation without a RANSAC stage. Training, benchmark evaluation, demo visualization, 3DMatch overlap preparation, and shared geometry utilities are represented. Several dataset, loss, and evaluation implementations were not excerpted; their nodes reflect documented capabilities and tree paths, while unverified connections are omitted.
Main entry points: regtr.py REGTR network · trainer.py Training runtime · demo.py Registration demo · modelnet.py ModelNet loader
Components 24
Data preparation
- 3DMatch and ModelNet
- ModelNet loader
src/data_loaders/modelnet.py - 3DMatch loader
src/data_loaders/threedmatch.py - Pair transforms
src/data_loaders/modelnet_transforms.py - Pair transforms
src/data_loaders/transforms.py - Overlap preparation
src/data_processing/compute_overlap_3dmatch.py
Registration model
- KPConv encoder
src/models/backbone_kpconv/kpconv.py - KPConv blocks
src/models/backbone_kpconv/kpconv_blocks.py - Kernel points
src/models/backbone_kpconv/kernels/kernel_points.py - Transformer layers
src/models/transformer/transformers.py - Position embeddings
src/models/transformer/position_embedding.py - REGTR network
src/models/regtr.py - Registration losses
src/models/losses
Training and evaluation
- Training runtime
src/trainer.py - Training entry
src/train.py - Registration demo
src/demo.py - 3DMatch evaluation
src/evaluate_3dmatch.py
Geometry utilities
- Torch SE(3)
src/utils/se3_torch.py - NumPy SE(3)
src/utils/se3_numpy.py - Torch Lie groups
src/cvhelpers/lie/torch - NumPy Lie groups
src/cvhelpers/lie/numpy
Visualization
- Point-cloud visualization
src/utils/viz.py - Visualization backend
src/cvhelpers/visualization
Other components
- Researcher
Connections 24
- Researcher → Training entry: starts training
src/train.py - Training entry → Training runtime: runs
src/train.py - Training runtime → REGTR network: trains model
src/trainer.py - Researcher → Registration demo: runs demo
src/demo.py - Registration demo → REGTR network: loads model
src/demo.py - Registration demo → Point-cloud visualization: visualizes result
src/demo.py - Researcher → 3DMatch evaluation: runs evaluation
- 3DMatch and ModelNet → ModelNet loader: provides samples
src/data_loaders/modelnet.py - 3DMatch and ModelNet → 3DMatch loader: provides samples
- ModelNet loader → Pair transforms: applies transforms
src/data_loaders/modelnet.py - 3DMatch loader → Pair transforms: applies transforms
src/data_loaders/transforms.py - Researcher → Overlap preparation: precomputes overlap
- REGTR network → KPConv encoder: encodes clouds
src/models/regtr.py - KPConv encoder → KPConv blocks: uses blocks
src/models/backbone_kpconv/kpconv.py - KPConv blocks → Kernel points: uses kernels
src/models/backbone_kpconv/kpconv_blocks.py - REGTR network → Transformer layers: attends features
src/models/regtr.py - REGTR network → Position embeddings: encodes positions
src/models/regtr.py - REGTR network → Registration losses: computes losses
src/models/regtr.py - Training runtime → ModelNet loader: consumes batches
- Training runtime → 3DMatch loader: consumes batches
- REGTR network → Torch SE(3): uses pose operations
src/models/regtr.py - Torch SE(3) → Torch Lie groups: uses Lie groups
src/utils/se3_torch.py - NumPy SE(3) → NumPy Lie groups: uses Lie groups
src/utils/se3_numpy.py - Point-cloud visualization → Visualization backend: renders objects
src/utils/viz.py