MonoLite

Explore lightweight practices for monocular 3D inspection

探索单目3D检测的轻量级实践

[中文][English]

Abstract摘要

Note that we are an engineering project, the code will be updated synchronously, currently in the early stages of the project, if you want to help, please check out our projects!

注意，我们是工程化项目，代码会同步更新，目前处于项目的早期阶段，如果你想提供帮助，请查阅我们的projects!

Activity活动

Design架构设计

我们将神经网络训练中最重要的部件分离了出来，而其他针对模型的操作，如训练/测试/评估/导出，则作为一种任务文件被不同的实验共用。

Experiment实验

Model	Dataset	info
MonoLite	Kitti

Torch info

Model	Input size (MB)	Params size (MB)	Total params	Total mult-adds
MonoLite	94.37	109.04	27,260,609	903.20

性能测试

*We used the BN layer, so a value of >=2 is recommended

Task	GPU(GB)	RAM(GB)	Batch size	Speed(it/s)
train	1.2	2.2	1
train	1.8	2.2	2
eval	2.2	2.0	1	43

Environment环境

虚拟环境

依据此pytorch_issue中的讨论，我们将虚拟环境迁移至miniforge

conda create -n monolite python=3.12

前置组件

系统	组件	下载URL	备注
windows	Visual Studio 2022	download	注意不同版本间的冲突
windows	Cmake	download	3.30.5
windows	MSbuild	通过VS2022下载
windows	MSVC	通过VS2022下载	手动添加至环境变量PATH

pip

pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt

Docker（暂不可用）

set DOCKER_BUILDKIT=0
docker build -t monolite .
docker run -d  --privileged=true --net host --name {any_name} --shm-size 4G --ulimit memlock=-1 --gpus=all -it -v C:\:/windows/ monolite:latest /bin/bash

Docker mirror

https://github.com/DaoCloud/public-image-mirror

Dataset数据集

Kitti

TODO

Pre-training model zoo预训练模型

Model	URL	Trainning log
MonoLite_Baseline	[百度网盘][谷歌网盘]	https://swanlab.cn/@Sail2Dream/monolite/overview

Inference推理

python tools\detect.py --cfg C:\workspace\github\monolite\experiment\monolite_YOLO11_centernet

Train训练

python tools\train.py --cfg C:\workspace\github\monolite\experiment\monolite_YOLO11_centernet

Train with your own Dataset自定义数据集训练

TODO

Eval评估

TODO

Export导出

ONNX

python tools\export_onnx.py --cfg C:\workspace\github\monolite\experiment\monolite_YOLO11_centernet

TensorRT

torchscript

python tools\export_ts.py --cfg C:\workspace\github\monolite\experiment\monolite_YOLO11_centernet

exported_program

python tools\export_ep.py --cfg C:\workspace\github\monolite\experiment\monolite_YOLO11_centernet

Torch_JIT

python tools\export_pt.py --cfg C:\workspace\github\monolite\experiment\monolite_YOLO11_centernet

Confirm致谢

我们衷心感谢所有为这个神经网络开源项目做出贡献的个人和组织。特别感谢以下贡献者：

type	name	url	title
CVPR 2021	MonoDLE	monodle github	Delving into Localization Errors for Monocular 3D Object Detection
3DV 2024	MonoLSS	monolss github	Learnable Sample Selection For Monocular 3D Detection
	TTFNet
community	kitti_object_vis	kitti_vis github	KITTI Object Visualization (Birdview, Volumetric LiDar point cloud )
community	mmdet3d	mmdet3d github	OpenMMLab's next-generation platform for general 3D object detection.
community	ultralytics	ultralytics github	YOLOv8/v11+v9/v10
community	netron	netron web	Visualizer for neural network, deep learning and machine learning models
community	mkdocs-material	mkdocs github	Documentation that simply works

正是这种协作和共享的精神，让开源项目得以蓬勃发展，并为科技进步做出贡献。我们期待未来有更多的合作和创新，共同推动人工智能领域的发展。

再次感谢每一位支持者，你们的贡献是无价的。