Bounding-box viewer for Jupyter

One afternoon, I needed to verify bounding box annotations that our labelers did for a bunch of receipt images. Instead of setting it up a separate job in AWS Sagemaker A2I, I decided to write a bbox viewer that I can use from Jupyter as a widget. Here’s a quick look of what it looks like:

This is not a new concept. There are already similar and full-fledged implementations like this, this, and this. This prototype however is intended for reviewing labeled receipt images where there could be several labels/words/tokens to verify. There are also minor QOL features like:

• Load the dataset from a csv file
• Add new columns to the csv file with text annotations (great way to add anecdotes to your data like “Missing amount”, “Skip”, etc.)
• Auto-save to csv as you’re traversing through the dataset
• Jump to row number
• Configure individual window sizes for the main image panel, text edit panel, etc.

I’m sharing this to get inputs on how to improve the tool. Feel free to ask questions, PR, or fork the repo:

GitHub - ismaelc/labelone