Folks in the Lean Startup up community talk a lot about the Minimum Viable Product, but I needed something way less than that. Something that needs to accomplish tasks that were core to the end goal I had in mind. So I came up with a list of things that I need this app nay, tool, to do.
One of the biggest things I struggled with when labelling large quantities of images from my phone was that fact that they were all similarly named and they all looked really similar. I need to be able to tell which images have been labelled and which ones haven’t.
Each image in an Object Detection dataset is accompanied by a file or JSON entry that contains coordinates and dimensions that represent where on the image the training algorithm needs to look for the area of interest. It’s not practical for a human to hand key in all of the coordinates and dimensions so the next thing I need is a way to allow users to pull up an unlabelled image and with a cursor draw a rectangle around the object they want the model to detect.
Once the rectangle has been drawn we need a away to associate that rectangle to a label, which will then be used by the training algorithm.
I would also like to be able to query images by label so that I can selectively choose which groups of images get added to my dataset.
Lastly, I need to be able to output all of my labelled data in a format that the training algorithms can ingest easily.
