search

yijingru / BBAVectors-Oriented-Object-Detection

[WACV2021] Oriented Object Detection in Aerial Images with Box Boundary-Aware Vectors
MIT License
462 stars 87 forks source link

All test outputs on the custom data set are HBB #108

Open HolleworldJIN opened 2 years ago

HolleworldJIN commented 2 years ago

Hi, Thank you for your works. I trained your network on custom dataset (the training dataset is about 200, only one object per image with one category), and I found that all the test outputs are HBB, but the traing image and testing image have rotated object. I checked the training dataset and it's annotation is correct. In addtion, I observed that the main part of loss is the 'wh' loss, about 3~5. Do you have any suggestions to fix the problem?

Some Info:

  1. 4 * GTX3080Ti
  2. batch_size=20
  3. epoch=200
  4. image shape is (width:2446, height:2048, channel:3)
yijingru commented 2 years ago

You may want to check ground-truth for Orientationalpha https://github.com/yijingru/BBAVectors-Oriented-Object-Detection/blob/7efd410f5f4ded94aca986aeb0bd1292235d1314/datasets/base.py#L166