Skip to content

A project demonstrating Lidar related AI solutions, including three GPU accelerated Lidar/camera DL networks (PointPillars, CenterPoint, BEVFusion) and the related libs (cuPCL, 3D SparseConvolution, YUV2RGB, cuOSD,).

License

Notifications You must be signed in to change notification settings

LinuxCup/Lidar_AI_Solution

 
 

Repository files navigation

Lidar AI Solution

This is a highly optimized solution for self-driving 3D-lidar repository. It does a great job of speeding up sparse convolution/CenterPoint/BEVFusion/OSD/Conversion.

title

Pipeline overview

pipeline

GetStart

$ git clone --recursive https://github.com/NVIDIA-AI-IOT/Lidar_AI_Solution
$ cd Lidar_AI_Solution
  • For each specific task please refer to the readme in the sub-folder.

3D Sparse Convolution

A tiny inference engine for 3d sparse convolutional networks using int8/fp16.

  • Tiny Engine: Tiny Lidar-Backbone inference engine independent of TensorRT.
  • Flexible: Build execution graph from ONNX.
  • Easy To Use: Simple interface and onnx export solution.
  • High Fidelity: Low accuracy drop on nuScenes validation.
  • Low Memory: 422MB@SCN FP16, 426MB@SCN INT8.
  • Compact: Based on the CUDA kernels and independent of cutlass.

CUDA BEVFusion

CUDA & TensorRT solution for BEVFusion inference, including:

  • Camera Encoder: ResNet50 and finetuned BEV pooling with TensorRT and onnx export solution.
  • Lidar Encoder: Tiny Lidar-Backbone inference independent of TensorRT and onnx export solution.
  • Feature Fusion: Camera & Lidar feature fuser with TensorRT and onnx export solution.
  • Pre/Postprocess: Interval precomputing, lidar voxelization, feature decoder with CUDA kernels.
  • Easy To Use: Preparation, inference, evaluation all in one to reproduce torch Impl accuracy.
  • PTQ: Quantization solutions for mmdet3d/spconv, Easy to understand.

CUDA CenterPoint

CUDA & TensorRT solution for CenterPoint inference, including:

  • Preprocess: Voxelization with CUDA kernel
  • Encoder: 3D backbone with NV spconv-scn and onnx export solution.
  • Neck & Header: RPN & CenterHead with TensorRT and onnx export solution.
  • Postprocess: Decode & NMS with CUDA kernel
  • Easy To Use: Preparation, inference, evaluation all in one to reproduce torch Impl accuracy.
  • QAT: Quantization solutions for traveller59/spconv, Easy to understand.

CUDA PointPillars

CUDA & TensorRT solution for pointpillars inference, including:

  • Preprocess: Voxelization & Feature Extending with CUDA kernel
  • Detector: 2.5D backbone with TensorRT and onnx export solution.
  • Postprocess: Parse bounding box, class type and direction
  • Easy To Use: Preparation, inference, evaluation all in one to reproduce torch Impl accuracy.

CUDA-V2XFusion

Training and inference solutions for V2XFusion.

  • Easy To Use: Provides easily reproducible solutions for training, quantization, and ONNX export.
  • Quantification friendly:PointPillars based backbone with pre-normalization which can reduce quantization error.
  • Feature Fusion: Camera & Lidar feature fuser and onnx export solution.
  • PTQ: Quantization solutions for V2XFusion, easy to understand.
  • Sparsity: 4:2 structural sparsity support.
  • Deepstream sample: Sample inference using CUDA, TensorRT/Triton in NVIDIA DeepStream SDK 7.0.

cuOSD(CUDA On-Screen Display Library)

Draw all elements using a single CUDA kernel.

  • Line: Plotting lines by interpolation(Nearest or Linear).
  • RotateBox: Supports drawn with different border colors and fill colors.
  • Circle: Supports drawn with different border colors and fill colors.
  • Rectangle: Supports drawn with different border colors and fill colors.
  • Text: Supports stb_truetype and pango-cairo backends, allowing fonts to be read via TTF or using font-family.
  • Arrow: Combination of arrows by 3 lines.
  • Point: Plotting points by interpolation(Nearest or Linear).
  • Clock: Time plotting based on text support

cuPCL(CUDA Point Cloud Library)

Provide several GPU accelerated Point Cloud operations with high accuracy and high performance at the same time: cuICP, cuFilter, cuSegmentation, cuOctree, cuCluster, cuNDT, Voxelization(incoming).

  • cuICP: CUDA accelerated iterative corresponding point vertex cloud(point-to-point) registration implementation.
  • cuFilter: Support CUDA accelerated features: PassThrough and VoxelGrid.
  • cuSegmentation: Support CUDA accelerated features: RandomSampleConsensus with a plane model.
  • cuOctree: Support CUDA accelerated features: Approximate Nearest Search and Radius Search.
  • cuCluster: Support CUDA accelerated features: Cluster based on the distance among points.
  • cuNDT: CUDA accelerated 3D Normal Distribution Transform registration implementation for point cloud data.

YUVToRGB(CUDA Conversion)

YUV to RGB conversion. Combine Resize/Padding/Conversion/Normalization into a single kernel function.

  • Most of the time, it can be bit-aligned with OpenCV.
    • It will give an exact result when the scaling factor is a rational number.
    • Better performance is usually achieved when the stride can divide by 4.
  • Supported Input Format:
    • NV12BlockLinear
    • NV12PitchLinear
    • YUV422Packed_YUYV
  • Supported Interpolation methods:
    • Nearest
    • Bilinear
  • Supported Output Data Type:
    • Uint8
    • Float32
    • Float16
  • Supported Output Layout:
    • CHW_RGB/BGR
    • HWC_RGB/BGR
    • CHW16/32/4/RGB/BGR for DLA input
  • Supported Features:
    • Resize
    • Padding
    • Conversion
    • Normalization

Thanks

This project makes use of a number of awesome open source libraries, including:

  • stb_image for PNG and JPEG support
  • pybind11 for seamless C++ / Python interop
  • and others! See the dependencies folder.

Many thanks to the authors of these brilliant projects!

About

A project demonstrating Lidar related AI solutions, including three GPU accelerated Lidar/camera DL networks (PointPillars, CenterPoint, BEVFusion) and the related libs (cuPCL, 3D SparseConvolution, YUV2RGB, cuOSD,).

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • C++ 51.9%
  • Python 21.9%
  • C 12.3%
  • Cuda 11.9%
  • Shell 1.1%
  • Makefile 0.6%
  • CMake 0.3%