Open bc-arl opened 4 months ago
@bc-arl Hi, any success finding an answer? We are also on the same boat and looking for a better alternate to MMDetection. Any success finding a better alternative?
I'm the guy above and no, I haven't managed to find anything better than MMDetection
MMdetection can be a bit cleaner if you mainly use it for training, and try to avoid their runtime as much as possible. I.e export to onnx and try to stick to vanilla onnxruntime. Does require writing preprocessing / postprocessing code though to mimic what the model is expecting as input and make sense of the outputs, theres some examples of people doing this on github for both mmpose and mmdetection
Here's an example for rtmpose https://github.com/Tau-J/rtmlib
P.S this approach can be a bit of a timesink initially, so I wouldn't do it if you just need to get something working quickly. But for a long term project, probably worth it.
@cravies Hey, thanks for the response and sharing some useful information. I am trying some luck with detectron2 as follows: https://github.com/facebookresearch/detectron2
Currently, it looks good. Regarding YoloX, It's still Apache 2.0 License so maybe it could be used for commercial use.
Out of curiosity, what do you think about the SuperGradients repository? I've used it for a while, and it seems well-structured, similar to MMDetection, where all training-related configurations can be specified in a single file. However, unlike MMDetection, it doesn’t require several large, intertwined repositories to work. However the downside is that Deci was bought by Nvidia some months ago and for now it is unclear if the repo will be updated in the future or if it will retain it's apache licence.
Trying to train some object detectors for a business. We need to be able to use a repo with a commercial license Just wondering if this is still state of the art open source yolo trainer as the repo hasn't been committed to in a year... We had been using MMDetection but it is a pain and we're looking to move off of it.