当前位置：首页 > news >正文

Siamrpn跟踪模型转RKNN模型

news 来源：原创 2025/6/15 3:28:36

基于Pysot（github仓库：Pysot），使用Siamrpn_alex模型进行模型转换，并在RK3588进行测试，实际测试FPS：15。

可访问我的github仓库进行复现（pysot_rknn）

Installation

you need to download rknn-tookit2,and install rknn-tookit2 on your ubuntu(x86-64,amd64)

export rknn and test

You need to download model from PySOT Model Zoo.

cd to rootPath(pysot_rknn)
python rknn/siamrpn_alex_dwxcorr.py --config experiments/siamrpn_alex_dwxcorr/config.yaml --snapshot experiments/siamrpn_alex_dwxcorr/model.pth --video demo/bag.avi

The file siamrpn_alex_dwxcorr.py can export rknn model, and test the result on your ubuntu(x86-64,amd64).

Detail about siamrpn_alex_dwxcorr model.
backbone of exemplar.

Target img:

input size: torch.Size([1, 3, 127, 127])
outputsize: torch.Size([1, 256, 6, 6])

backbone of instance.

Original img:

input size: torch.Size([1, 3, 287, 287])
outputsize: torch.Size([1, 256, 26, 26])

rpn Head

rpn head input size: [z_f,x_f]

z_f shape: torch.Size([1, 256, 6, 6])
x_f shape: torch.Size([1, 256, 26, 26])
output size: output[[cls],[loc]]
cls shape: (1, 10, 21, 21)
loc shape: (1, 20, 21, 21)

z_f is the output of backbone of exemplar, x_f is the output of backbone of instance

RUN the demo on RK3588

Install rknn-toolkit-lite2 on your rk device, I run the demo on RK3588.
Copy folders and files on the projict to your rk device, and your need to maintain the structure of the project.