search

weecology / DeepForest-pytorch

Pytorch implementation of the deepforest model for tree crown RGB detection.
MIT License
16 stars 9 forks source link

Prediction accuracy #80

Closed Klara8A closed 3 years ago

Klara8A commented 3 years ago

Hi! I started using DeepForest in April when it was still running with tensorflow and now I have started testing it with the new version based on pytorch. I have noticed that the predictions are different. Being those of the new version less accurate and with more "errors". I have tried it with different patch sizes, and different iou. So I was wondering if this is due to the changes that have been made and if this is going to improve in the following releases ?

I have not yet tested it after training, but in general I feel that the results of the untrained model are less accurate than before.

bw4sz commented 3 years ago

A couple users had reported similar issues. I haven't been able to reproduce at all, the benchmark score is within a few percent, see the other issues. Can you paste images (before and after) and the code you used? I'm considering training a new model, but I need to understand where differences might be, since the training data is identical.

bw4sz commented 3 years ago

also, can we move this to the deepforest repo, this repo is deprecated as we merge the two.