Open matt-sharp opened 3 years ago
have a look at the excellent updates for automatic builds with deps for linux and win the guys have been pushing out lately. pythonic voodoo only gets you so far with realtime inferencing but we are spoiled to have darknet build scripts automated
I'm in the exact same position as you. haviduck has a good point with the scripts that will do most of the work for you but I was wondering if you ever figured anything out? I've been trying to see if loading darknet directly like how it's done in darknet_video.py
but I haven't gotten it working yet.
if you guys can describe a little bit more we can certainly help. to load darknet.py you also need to compile the dlls as described in the readme, then place it somewhere you can refererence in darknet.py (search for the lines where he defines SO files like dlls). additionally you need to place some of the opencv dlls either next to darknet or have them in path. same goes for cuda. easily scripted and automated.
I have a YOLOv4 model trained using Darknet which I've tested for inference using OpenCV with the DNN module and it works great. It's fast, accurate, flexible and the inference script is written in Python so it's easy for me to maintain. The only complexity is that it requires OpenCV to be built from source to get the best performance.
My end-user is restricted in the software they can install and building OpenCV from source adds in too much complexity. Can anyone recommend an inference engine that is available from PyPi without any additional dependencies?