Apologies for asking again but I still have not received any usable advice regarding this issue.
The logitech c920 is the recommended camera for use with the jetson-inference program. There are no instructions on how to capture images with it to make new datasets for training in digits.
Please advise the recommended image capture approach including capture program, RGB formats etc.
Is such a pipeline even needed? @dusty_nv 's page only cites needing jpeg, tiff, bmp or png (with reference to retraining an existing network with images).
Does pixel format have any bearing on running inference on a newly trained dataset, given that you can train a network on random internet images and use any camera to run inference?
I think the pixel format doesn’t make too much different on it if they are all in the RGB color space.
Suppose a model should be robust enough on the visually identical images.
But you are afraid of the capturing process, color conversion, …, will cause a slight difference in the model.
You can use the same pipeline to save image for training.
Ok thank you. I am having trouble getting a trained network to function well, hence the questions. The test data set returns 100% accuracy, but with the exact same camera setup the inference is really struggling. I wanted to make sure that the colorspace was not a factor, and I think the message I am getting from you guys is it probably doesn’t have an impact, but is worth testing to see if it makes a difference.
Perhaps it could just be some quirk of my training settings, I am not sure. Anyway thank you for your help.
It sounds like an overfitting issue to me.
Accuracy should not reach 100%. Your model may try to memorize the input data rather than learn to classify.
The accuracy was on a separate test set, images not used for training or validation. Doesn’t this rule out overfitting?