2

I'm not sure if the Tensorflow ObjectDetection API automatically normalizes the input images (my own dataset). It seems to have an option called 'NormalizeImage' in the DataAugmentations. So far, I haven't specified it, and my models are doing reasonably well. Am I missing image normalization, or does Tensorflow do it automatically for me, or is it just not needed for this Object Detection API?

My models have used Faster RCNN and RetinaNet so far.

sabreshack
  • 558
  • 1
  • 4
  • 11
  • same question.. any news on this ? – rok May 13 '20 at 12:17
  • I think I answered it when I came across the same problem. https://stackoverflow.com/questions/61778254/normalize-input-for-tensorflow-object-detection-api/73020511#73020511 – Burschken Jul 18 '22 at 10:11

2 Answers2

0

Depending on what configuration you specified it uses an image resizer to normalize your dataset.

image_resizer { fixed_shape_resizer { height: 300 width: 300 } }

This will either downsample or upsample your images using bilinear interpolation.

GuyTraveler
  • 148
  • 7
  • 2
    I don't this that's the answer @sabreshack is looking for. Normalize meaning, kind of histogram equalization not resize. – Abhay Doke Mar 26 '20 at 22:41
0

If someone comes across this question.

Short answer: the features extractor that is defined in the training config is probably doing the preprocessing. If you are interested in a little more details, I answered it in a similar question. But when I added the link to the question my answer was transformed to a comment. So check out the comments of the original question.

Burschken
  • 77
  • 11