MiDaS(Multiple Depth Estimation Accuracy with Single Network)单目深度估计,部署 rk3588。
模型转换和仿真测试参考 onnx转rknn。
由于模型文件比较大,无法和代码一起上传。rknn模型文件
1)编译
cd examples/rknn_MiDaSDepth
bash build-linux_RK3588.sh
2)运行
cd install/rknn_MIDASDepth_demo_Linux
./rknn_MIDASDepth_demo
注意:修改模型、测试图像、保存图像的路径,修改文件为src下的main.cc
int main(int argc, char **argv)
{
char model_path[256] = "/home/firefly/zhangqian/rknn/examples/rknn_MiDaSDepth/model/RK3588/midas_v21_384.rknn";
char image_path[256] = "/home/firefly/zhangqian/rknn/examples/rknn_MiDaSDepth/test.jpg";
char save_image_path[256] = "/home/firefly/zhangqian/rknn/examples/rknn_MiDaSDepth/test_result.jpg";
detect(model_path, image_path, save_image_path);
return 0;
}
板端时耗也贴出供参考:模型midas_v21_384,模型输入分辨率384x224, rk3588芯片。