gouthamreddykotapalle / image-dataset-augmentor Goto Github PK

I could see that our library is considerable slow when compared to the python library. I was wondering how we will handle the comparisions 😅?

Each manipulation function is now handling the randomness of the operation using min and max facrtors. So as per the current design, each of the output images go through all the operations in the pipeline. (for example, when probability is 1).

But now, when each manipulation function takes an additional probability parameter, how will this effect our output and what design will we be following here?

Capability to read a batch of images or a single image

Using the jpeg.h and jpeg.cpp wrapper code for simplicity since it is a wrapper written in STL around the jpeglib.h data structures that use c-style representations. This is a wrapper library around the commonly used licensed libjpeg/jpeglib.h library for reading and writing jpeg files.

We can use the image class mentioned in this library for the Image objects in our Augmentor class. There is a sample shrink/expand function as well in the main.cpp file. We can take inspiration from this function and build the other image manipulation APIs.

@mingen-pan Need your help resolving the bug. The code is in goutham_dev branch.

So when I have invert probability=0.1 and have given only 2 images in my pipeline (batch images). The code fails to run and errors out. But it works with probability=1.0

https://github.com/Gouthamkreddy1234/Image-Augmentation-Library/blob/95e10a3dc0f3981ff248e5e97295bfdd7c9a9228/Augmentor.cpp#L69

Besides the null ptr issue, I realize the Error: Empty input file is caused by this command. Don't know why.

Originally posted by @mingen-pan in #3 (comment)

Need to complete the random point case in zoom. Now handling only center.

Probability - Batch level
Randomness range - Image level

We we be passing these two parameters for each function for example -

Augmentor.rotate(probability=0.5, min_rotate=30, max_rotate=60)

1. Multi-threading support
2. Pushdown optimzation (lazy evaluations)