北京地平线信息技术有限公司用户664
DOSOD Deployment to RDK X5
输入“/”快速插入内容
DOSOD Deployment to RDK X5
DOSOD Deployment to RDK X5
用户664用户664用户1476用户1476
用户5159用户51592025年11月27日修改
1. Introduction
✏️
Instructions
•
If you want to deploy directly , please jump directly to "4. Board-side deployment"
•
If you want to do it yourself , please refer to the "do it yourself" of each step
•
If you want to do it yourself , please pay attention to the environment, the experimental host system is Ubuntu20.04 (above can be), the host has docker
This article mainly introduces how to deploy the open vocabulary model Decoupled Open-Set Object Detector (dosod) to RDK X5, the process is as follows
画板
What is Open Vocabulary Object Detection?
The goal of Open Vocabulary Object Detection is to enable machines to recognize and locate new categories of objects in images that do not appear in the training set. Read this article for more information.
What is DOSOD?
DOSOD is an open vocabulary model developed by D-Robotics to support real-time Open Vocabulary Object Detection tasks in robot systems. For details, please refer to GitHub , welcome everyone to star
What is RDK X5?
2. Model export
Model export mainly converts pth files to onnx files. "2.1 Download Already Available" has put the exported files here for direct use. If you want to do it yourself, please refer to "2.2 Do It Yourself" for exporting onnx files
2.1 Download already available
The following is the exported model, which can be used directly.
DOSOD | Train set | Size | AP50(Zero-shot Inference on COCO) | PTH link | onnx link | bin(int16) link | bin(int16+int8) link |
dosod-s | O365+GoldG | 640 | 51.0 | ||||
dosod-m | O365+GoldG | 640 | 57.1 | ||||
dosod-l | O365+GoldG | 640 | 60.5 |
2.2 Do it yourself (optional)
💡
Please refer to GitHub for DOSOD training. This article mainly introduces exporting from onnx, model conversion, and how to run on the X5 board
Prepare the environment
🍞
For quick export, this article uses the docker environment
Building mirroring
代码块
# build mirror
docker build -f Dockerfile -t d_robotics/dosod:v1 .
Start container
代码块
# start docker
docker run -itd \
--name dosod \
--shm-size 128G \
--gpus all \
--privileged=true \
-w /root \
d_robotics/dosod:v1
# enter the container
docker exec -it dosod bash