Scaling Inquiry: Adapting Image Values for Model Pretraining

Reflection Connection
1

Hello @team

In the Patch the Planet dataset, the provided volumes have values that far exceed the range of 0 - 255 typically found in standard images such as PNGs. However, the images provided in the Reflection Connection challenge are PNGs and therefore have values ranging from 0 to 255. My question is, how did you manage to bring the images within the 0 to 255 range? Did you apply a min-max scaler at the image level or did you use custom bounds? In the latter case, could you please provide us with the custom bounds so that we can reproduce the preprocessing during model pretraining?

2

Hi @baptiste.u

We used the min-max scaler to normalize each image pixel values (both for pre-training and inference) to range 0-1 with the following snippet:

im = cv2.normalize(np.array(im), None, 0.0, 1.0, cv2.NORM_MINMAX, cv2.CV_32FC1)

Images provided for Reflection Connection challenge were generated from wide-range seismic with normalization to standard PNG range at image level, so we believe that such approach should guarantee maximum consistency between pre-training and downstream data.

Onward Team

1 Like