Open BasisPoint opened 5 years ago
In our conference version, we use a simple ensemble method by doing NMS on the concatenation of predictions (before NMS) from multiple models. For example, if you have two models both output 300 detections before NMS, I just combine them together (600 detections) and do NMS on these 600 detections.
A better ensemble method for two stage detectors is to combine proposals from both models, and use this new set of proposals to generate detections from both models and take the average (because proposals are the same, so you can average the results directly).
I have trained several faster rcnn models with different backbones, then I want to ensemble these models for a better performance, but I am totally confused about how to do this for object detection, can you give me some suggestion or reference for this? Thank you anyway.