Open jordi-bird opened 7 years ago
@jordi-bird As far as I know the YOLO pre-trained weights are learned without any image pre-processing other than re-scaling the data (with 1./255). This is actually the case for many nets that use batch normalization.
Hello, I saw your work on the mcv-m5 project on github and I feel very good, but there is only code on github, no data and corresponding keras development kit. Can you pass a complete data such as relevant data and keras development kit? thank you very much!
In order to solve the YOLO preprocessing issues in our implementation, we wonder if the imagenet weights employed were created applying pre-processing. If that's the case, we would like confirm if it's the one detailed in data_loader.py as 'imagenet' or any more techniques were used.
Thanks for your time.