search

ubc-vision / IterativeSG

19 stars 2 forks source link

Performance of Pre-Trained DETR #3

Closed ShunchiZhang closed 1 year ago

ShunchiZhang commented 1 year ago

Hi Dr. Khandelwal @siddheshk,

Did you test the performance (e.g.: Average Precision) of your pre-trained DETR on Visual Genome dataset? Or would you mind giving me some instructions on how to do so using your codebase?

Thanks for taking your time!

siddheshk commented 1 year ago

I think this was the performance of the pre-trained DETR on Visual Genome detection.

AP AP50 AP75 APs APm APl
17.303 31.122 16.506 5.511 13.066 23.519

You can train DETR if needed using this codebase. You will need to use the following flags when running the training code. MODEL.META_ARCHITECTURE 'Detr' MODEL.DETR.NAME 'DETR' MODEL.DETR.TRANSFORMER 'Transformer' MODEL.DETR.CRITERION 'SetCriterion' MODEL.DETR.MATCHER 'HungarianMatcher' SOLVER.IMS_PER_BATCH '16' SOLVER.STEPS '(369600,)' SOLVER.MAX_ITER '554400' The learning rate and steps are taken from the original DETR implementation, which can be found here.