WarpConvNet
A high-performance 3D deep-learning library that brings sparse voxel convolution, point operations, and attention into one CUDA-native system.
Libraries and models for 3D perception
A curated index of the 3D learning stack, connecting software and neural operators to geometric features, registration, spatio-temporal models, and open-vocabulary scene understanding.
- Neural network libraries
- 3D representation learning
- 4D perception
- Registration
- Scene understanding
Implementations 02
The systems behind the work.
Core libraries for efficient 3D learning.
-
-
MinkowskiEngine
The auto-differentiation library for high-dimensional sparse tensors that introduced generalized sparse convolution as a practical neural-network operator.
Research lineage 07
From operators to perception systems.
A chronological index of papers and the dissertation.
-
4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks
Generalized sparse convolution and direct 4D spatio-temporal perception over sequences of 3D scans.
-
Fully Convolutional Geometric Features
A sparse fully convolutional network that extracts compact geometric features from complete point clouds in one pass.
-
High dimensional Convolutional Neural Networks for 3D Perception
The full arc from sparse tensor networks and 4D segmentation to geometric representation learning and high-dimensional registration.
-
High-dimensional Convolutional Networks for Geometric Pattern Recognition
Sparse convolutional networks operating in spaces from four to 32 dimensions to recognize correspondence geometry.
-
Deep Global Registration
FCGF features and a 6D sparse convolutional network inside an end-to-end system for robust point-cloud registration.
-
Mosaic3D: Foundation Dataset and Model for Open-Vocabulary 3D Segmentation
A sparse 3D encoder trained on 5.6 million mask-text pairs for open-vocabulary semantic and instance segmentation.
-
SpaCeFormer: Real-Time Open-Vocabulary 3D Instance Segmentation Without Proposals
A hierarchical sparse-voxel U-Net that combines sparse-convolution shortcuts with spatial-window and space-curve attention.