Question about IR Transformation in Preprocessing
Hey @wiktorlazarski
I'm assuming that the original photos from the datasets were in a simple RGB scale and the infrared transformation occurred as a preprocessing step before feeding the model for training. If that's the case, I was wondering:
- How can this IR transformation be simulated using pure digital signal processing (DSP) techniques or similar?
- Did you use a specific algorithm or library to achieve this transformation?
I'm working on adapting a similar pipeline and would appreciate any guidance or references to reproduce this step.
Thanks in advance for your insights!
Hey @gkont ,
Thank you for your interest in open-iris project and your question.
The input to a neural network when training/making inference is/was always IR image (single channel image). To utilise pretrained on ImageNet model while training the model, we decided that we will duplicate two times a single channel IR image. That gave us a tensor with 3 channels, so that we can input this image to neural network as if it was RGB image. Here is the preprocessing source code that does entire preprocessing of input IR image https://github.com/worldcoin/open-iris/blob/main/src/iris/nodes/segmentation/multilabel_segmentation_interface.py#L41.
Hope that helps answering your question!
Best regards,
Wiktor