object_level_visual_reasoning icon indicating copy to clipboard operation
object_level_visual_reasoning copied to clipboard

how to train mask-r-cnn

Open JackSparrow3 opened this issue 4 years ago • 1 comments

Hi Fabien, thank you for your works! the paper doesn't mention how to train a mask-r-cnn since VLOG dataset does has any mask annotation, i wonder if your team has annotate the mask by yourself or use other open datasets to train mask-r-cnn? thanks again!

JackSparrow3 avatar Aug 20 '19 11:08 JackSparrow3

Hi @JackSparrow3 Thanks :) You can find the annotations here: https://fabienbaradel.github.io/masks_data/ We have used MaskRCNN pretrained on COCO. We did not train MaskRCNN on our datasets (VLOG, Something Something, EPIC) because we do not have access to object labels. We have only used the prediction made by Maskrcnn. I hope it is clear now. In case you want to train MaskRCNN you can have a look at this repo: https://github.com/facebookresearch/maskrcnn-benchmark

Fabien

fabienbaradel avatar Sep 10 '19 20:09 fabienbaradel