CPU Generated Birds Eye View from KITTI Data

voxelize

A CUDA C++ kernel that converts raw LiDAR point clouds into the pillar-voxel representation consumed by 3D object detection networks.

---

Background

This project takes point clouds and turns them into 3D pillars in a BEV grid. This allows for 3D space to be processed by a normal CNN the same as a 2D image would be.

---

Evaluation

1. Benchmark table comparing this kernel against OpenPCDet's voxel_generator at four point cloud densities on an RTX 5060 Ti:

Points	This kernel (ms)	OpenPCDet (ms)	Speedup
25k	1.050 ± 0.012	4.327 ± 0.003	4.12x
50k	1.111 ± 0.010	9.088 ± 0.006	8.18x
100k	1.150 ± 0.010	23.497 ± 0.033	20.43x
130k	1.207 ± 0.011	34.918 ± 0.009	28.92x

---

Repository Structure

voxelize/
├── csrc/
│   ├── voxelize.cu
│   ├── voxelize.h
│   └── voxelize_ext.cpp
├── voxelize/
│   ├── __init__.py
│   └── cpu_reference.py
├── tests/
│   ├── test_correctness.py
│   └── test_benchmark.py
├── benchmarks/
│   ├── benchmark.py
│   ├── profile_kernel.py
│   └── baseline.json
├── scripts/
│   └── explore_kitti.py
├── .github/
│   └── workflows/
│       └── ci.yml
├── setup.py
├── Makefile
└── README.md

---

Hardware Requirements

This runs entirely on an RTX 5060 Ti 16GB. Only kernel execution and benchmarking is done here, no model training.

Resource	Requirement
GPU	Any NVIDIA GPU
VRAM	2GB minimum
CUDA Toolkit	12.x
Python	3.10+
PyTorch	2.x

---

Sources

PointPillars (Lang et al., CVPR 2019)