Labeling Functionality
FINISHED
- Task 1: Could change the name of wrongly detected classes
- Task 1: Can add bounding boxes around not detected objects
TO-DO
- Task 1: Support for Darknet
Need of Labeling Functionality.
As mentioned in my previous post , this tool can also be used to create datasets. And as we all know Deep Learning has a never ending quench for data , tools like DetectionSuite are needed , which use trained deeplearning models to inturn generate datasets for deeplearning. So what did I do ? ……….wait , I’ll tell. In this world where people are working their brains off to increase their model’s accuracy by even 1% , it would be unfair towards them if the data itself is not accurate. Although I could not make it 100% correct, effort was made to make the life of the person little easier , who is using this tool to generate training data.
How ?
- While using this tool if you think the object detected is not properly enclosed i.e. over/under enclosed then you could change the bounding boxes accordingly to correct it.
- If you think a car has been labeled as a cell phone , you could change that as well just by clicking on the classname.
- If the object itself is not detected , you could draw bounding boxes around multiple such detections and add their class names yourself.
Tool in action.
Problems faced
The main issue I had while developing was with Qt , it was working asynchronously. This could be a boon in many situations , but in my case I needed my program to pause execution until it receives some input from the user using the GUI . Although we could write a while loop like in case of pause button in playback functionality , I wanted to do something little more efficient. On surfing the internet for a while , I found Qt has some nice built in librarie functions for the same. Although the fix was small it plucked my nerves for a while.
Suggestions
Currently the datasets are being stored in one format only. Support needs to be added to store dump them in other formats like COCO, Pascal etc. And as of now this feature does not have a GUI of its own. But once after adding the support, it’s own workspace would be very helpful.