Ensemble detectron2 weights #3559
Description
I have been experimenting with the different model weights from the model_zoo, and each of them shine in different ways. For instance, I have realized keypoint is able to detect more humans/persons than the other model types, but that's all it can do since it's built for that. Is there a way to load pretrained weights of keypoint and, say, weights of PointRend to train a single model?
More like performing weight ensembling to increase the performance of the prediction, or more specifically to detect more objects from a single input sample.
Thank you.