Inspiration

We’re getting old, and with the cost of living as high as it is, we wanted a fun (and free!) way to get some exercise. That brought back memories of playing Just Dance on the Wii. But modern consoles and games aren’t cheap, so we decided to bring the experience back ourselves: the classics, with no expensive hardware, and ideally the ability to dance along to any Just Dance video on YouTube.

What it does

Maybe Boogie turns almost any Just Dance video on YouTube into a playable dance game. Paste a link, and our servers process the video automatically. They analyse the choreography, extract the dancer’s poses and sync them to the music to build a playable track.

Don’t want to wait? Browse our library of ready made tracks and jump straight in. Using a webcam, laptop camera or even your phone, Maybe Boogie tracks your movements in real time and scores you on how closely you match the choreography.

No Wii, no console, no motion controllers. Just your device, a camera and your best moves. Have you still got what it takes to beat your childhood high scores?

How we built it

We built Maybe Boogie as a Django web app so we could get the groundwork done quickly and spend our time on the interesting parts. We started with login, logout and admin pages, authenticated user pages and a relational SQLite database. While that was in development, we drew our own assets and animations, and even made a custom font to give the whole thing that hacked-together feel. (Anatomy and proportions don’t matter in the artwork. They only matter in the dancing.)

Next came YouTube links, which were trickier than we expected. When you paste a link, a regex strips it down to just the video ID, and we download the video at 480p. Because we wanted a multi-user site, all of this runs in a background thread with robust error handling, so one person’s video doesn’t hold everyone else up.

Then the exciting part: cameras! Camera access in JavaScript is fairly simple, and Google’s MediaPipe has a web model that runs directly in the browser. It’s light enough to run on phones, which was one of our goals from the start. Running it alongside the video and a progress bar wasn’t too tricky either. Unfortunately, that’s where our luck ran out, because next we had to figure out the maths.

Pose models provide (x, y) coordinates for each joint. That’s great, but it doesn’t tell you whether someone is busting the right move. So we split the pose into limb segments and, with some fun mechanics (can you tell we hated GCSE physics?), turned each one into a normalised unit vector: the direction that limb is pointing. Because only the direction counts, it doesn’t matter how tall you are or how far you stand from the camera. Ta-da! We then added weightings to the score so that your arm movements are worth more points than your leg and torso movements, replicating the "I'll just wave my controller around and hope" movements we all know and love!

Want a backstage look?: https://youtu.be/tTRUlilkQCw

Challenges we ran into

Pose models are trained on real people… not Just Dance people! The game’s stylised dancers often weren’t recognised as humans at all, and the busy backgrounds didn’t help. We spent a long time fighting with YOLO and bounding boxes, and eventually added a tolerance window around each dancer so the model focuses on them rather than everything happening behind them.

Accomplishments that we’re proud of

Every asset is hand-drawn, and together they have an uncanny-valley feel that makes the experience one of a kind. We also put real thought into our scoring labels, brainstorming synonyms just far enough from the originals to feel familiar without copying them. We figured out how to get some pretty accurate pose comparisons which was cool.

What we learned

Video processing is much harder than we expected, and Just Dance dancers don’t count as people, at least not to a pose model. We also finally learned how to use uv properly. Amy picked up some basic animation and learned how to design and export a font. Char experimented with Vultr, discovered that caching with Django is a pain, and got to grips with basic vector maths (“normalising a vector into a unit vector… blah blah blah”).

What’s next for Maybe Boogie

Multiplayer! Live leaderboards! Stress testing (of us and the servers)! Auto-trimming videos with awkward intro or outro cards…

How it fits "Composition"

Maybe Boogie is our take on composition: it brings together art, music and tech. Every asset, animation and even the font is hand-drawn by us. The music and choreography come from whichever video you choose. And the tech (pose extraction, real-time computer vision and some vector maths) is what turns the two into something you can play.

Hackiest Hack

Maybe Boogie is held together with string, and we’re proud of it. We took videos of cartoon dancers and fed them to pose models that refuse to believe Just Dance dancers are people (they deserve rights too!). Then we fought YOLO and bounding boxes until they gave in, and used physics we swore we’d left behind at GCSE. All of it runs on a Django app with background threads, a regex that strips YouTube links down to a bare ID, and a hand-drawn custom font designed to look as hacked-together as the project really is. And caching? Caching is a nightmare hellhole we never want to see again. Who needs a custom UI when you can just slap graphics right on top of a video? What could go wrong?

Could we have just bought a Wii? Yes. Did we rebuild one out of a webcam, YouTube and spite instead? Also yes.

Accenture

Money is tight, and dance games ask a lot of it. First you buy the console, then the game, then the motion controllers, and you buy it all again when the next version comes out. Older versions often never make it to newer consoles at all. People accept this as just how gaming works, so we didn’t.

Maybe Boogie replaces all of that with things almost everyone already has: a device with a camera, and YouTube, where years of dance routines are already freely available. AI does the job the hardware used to do. Computer vision extracts the choreography from the video, and a pose model running in your browser tracks your movements in real time and scores you. No console, no controllers, no paywall. Just paste a link and dance.

Best Use of Vultr

Our pose-processing pipeline ran on our own laptops, and it was slow: one video took 15 minutes to process, leaving players a long wait before they could dance. Vultr removed that bottleneck. We moved video preprocessing (downloading, YOLO detection, and pose extraction) to Vultr, reducing processing time to 2 minutes. Because it runs in the cloud and not on someone’s laptop, several users can submit videos at once without everything grinding to a halt. We specifically chose the cloud GPU option so that we could have fast processing for our players. Less waiting, more dancing.

Share this project:

Updates

Submission history