hashirzahir / transfer-learning Goto Github PK

Can use ML or CV based image augmentation techniques to add more variability to the image data set to allow for a more accurate trained model.

There seems to be a bug in the source code where after more than 60% or so of images labelled, previously ignored (when the ignore button is pressed) or labelled images keep appearing (without their annotation as well) and causes difficulty to finish the last 40% of images (eg: May require 100 labelling attempts to finish an image set with 50 images). Still have to ensure that all users in network receive different images at the same time.

Feature was intended to allow for reviewing by other people to verify that the bounding box was accurate but the bounding box does not appear either, making it seem like a new image to label for the user. For new review feature issue, look at #3.

Main source file located at index.php and getNewImage.php

If an image was ignored a certain number of times (eg: ignore button was pressed twice for the same image by different people at different times), then the image should no longer be put up for display to the users. Should also have a way to remove from the train set data.

Add a separate tab or button that allows the user to switch to "review mode" where he can help to verify annotations made by other users or themselves to prevent bad data from being fed to the ML model for training.

Also need to ensure that images are not repeatedly reviewed (eg: image should not appear on both the review and annotate pages of the website once reviewed by a user in the "review tab")

Current ssdlite_mobilenet_v2 model from the tensorflow model zoo works fast on embedded systems (aka Jetson Nano) but has low accuracy. Move to a new ML model for transfer learning or other methods