Using cupy and onnx-runtime-gpu to boost the performance on Robust Video Matting[link] for onnx.
On my i7-9700k and RTX3060 win10 computer, two video matting only takes 40% cpu and 60% gpu.
2022/09/07 add two srcipts for inference on cpu.
Time consuming:
~30S for converting a 30s video on HDD.
~22s for converting a 30s video on SSD.
2min for 30s video on an i9 cpu.
Gpu:
FFmpeg
Opencv-python
Cuda and cudnn
Numpy and cupy(For windows, you need to install the visual studio first, then run or re-run the cuda installer)
Onnxruntime-gpu
tqdm
Note:The package version depends on your computer.
Cpu:
onnxruntime
tqdm
opencv-python
numpy
Note:You'd better create a new env for install.
1.Download my project zip or git clone.
2.put your test video(avi) in the same directory and rename input.avi.
3.run onnx_gpu_infer.py for FP16 inference or onnx_gpu_infer_fp32.py for FP32 inference(it takes twice the time)
or run onnx_cpu_infer.py for FP16 inference or onnx_cpu_infer_fp32.py for FP32 inference
cpu is better on fp32 than fp16
4.Then look at the com.avi result.(If it's 0k means have some errors on your opencv)
For further usage, just change the code whatever you like(e.g. background color, video format).