search

cvlab-stonybrook / LearningToCountEverything

MIT License
361 stars 75 forks source link

Batch Size #18

Closed 1171000410 closed 2 years ago

1171000410 commented 2 years ago

Hello, We noticed the input of FamNet is the whole image and batch size equals to 1. Have you tried random crop and increase the batch size?

Best wishes! Thank you for your answers in advance!

Viresh-R commented 2 years ago

Hey, We did not use random crops with batchsize > 1 in the paper since the crop might not contain one or more of the exemplars. In order to use batchsize > 1, you'll need to address this case. One easy way of doing so would be to not use ROI pooling for those exemplars which go outside the crop, and computing the Resnet features for such exemplars independently.