search

poodarchu / learn_aug_for_object_detection.numpy

Learning Data Augmentation Strategies for Object Detection in numpy
55 stars 8 forks source link

A bug In func : _shear_bbox #2

Closed heartInsert closed 5 years ago

heartInsert commented 5 years ago

In function shear_with_bboxes=> _shear_bbox : your return 3 additional values bbox[4], bbox[5], bbox[6] compare to the orignal tesorflow_version line 1029, return tf.stack([min_y, min_x, max_y, max_x])

Question 1: in the orignal tesorflow_version , bboxes input format should [ymin,xmin,ymax,xmax],there is no bbox[5] and bbox[6], do you change the input format in numpy_version ? Questin 2: I saw a new policy named policy_custom, will it worked better than others ? The policy which I used is policy_custom, bboxes format is [ymin,xmin,ymax,xmax]. Thanks for your reply

poodarchu commented 5 years ago
  1. my dataloader need extra 3 values for a bbox, you can remove that.
  2. default policies are so strong that they'll cause performance decrease greatly. so I add a light augmentation version, it still doesn't work, but only with less performance decrease compared to original policies.