I Just Wanted to Look at My Own Images
A companion tool for Ultralytics users. Their platform exports datasets as NDJSON, so I built something that turns that back into images and labels you can actually open and browse locally.
I want to start by saying I love the Ultralytics platform. I use it all the time, and it's genuinely one of the best tools out there for vision work. So nothing here is a complaint about them. It's the opposite, really.
When you download a dataset from the platform, it comes to you as NDJSON. Inside their ecosystem that makes complete sense. But the moment I wanted to step outside it, just to look at my own images, check what my train/val split actually looked like, or see how things sat in plain folders on my machine, I couldn't. It's one file, not a dataset you can open and browse.
That was my whole itch. I kept wanting to do the most basic thing, open my dataset and look at it, and there was no easy way to do it locally.
So I went looking for a tool that would just turn the NDJSON back into images and labels. And I couldn't find one. Nothing out there did this. That genuinely surprised me, because it felt like such a normal thing to want.
So I built the YOLO NDJSON Converter. You hand it your NDJSON export, and it gives you your dataset back the way you'd expect, images you can open, splits you can see, in the format you actually want to train with, and not just YOLO. It converts to COCO, Pascal VOC, and every YOLO version too, so your data isn't locked to one place.
I think of it as a companion to Ultralytics, not a replacement for anything. It sits alongside the platform and lets your dataset travel further, out to your own machine, other frameworks, wherever you need it. If it saves someone time, that's the whole point.
And that part has been the nicest surprise. I built this for myself, but I've watched a bunch of other people hit the exact same wall and reach for it too. Turns out I wasn't the only one who just wanted to look at their own images.
If you've run into the same thing, it's here, and it's open source.