dance moves Computer Vision Project

alsuren

Updated 2 years ago

39

views

0

downloads
Classes (59)
balboa-close-embrace
balboa-closed
bluesclose
bow
celebration
charleston-sbs
circle-3-east
circle-4-east
circle-6-east
closedtoclosetransition
h2h-1-east
h2h-1-west
h2hjazz
h2hpassby-2-east
h2hpassbyfrombehind-1-west
jazz
jazz-east
jazz-open-east
jostling-7-west
leadout-2-south
lift
lindy-circle-4
lindy-circle-5
lindy-circle-7
lindy-circle-8
lunge-southeast
minidip-1-east
minidip-4-east
outsideturnpassby-4-west
passby-1-east
passby-1-west
passby-2-east
passby-4-east
push-4-east
solo-f
solo-m
stupidpretzelthingie-4-west
sugar-4-east
swing-close
swing-close-embrace
swingout-1
swingout-1-east
swingout-1-southeast
swingout-2-east
swingout-3-east
swingout-4
swingout-4-south
swingout-5-east
swingout-5-southeast
swingout-7
swingout-8-east
swingout-8-south
swingout-8-southeast
tommy-5-east
tuck-1-east
tuck-2-west
tuck-4-east
tuck-5-east
tuck-turn-5
Description

An attempt at tagging dance moves so that I can train a model to recognise them.

Classes

Generally classes are named $move-$count-$direction (for example swingout-4-east) where:

$move is the name of a common move that is like/is structured like what they are doing (it isn't always what they're actually doing). If it's a common move then it will have a short $move, and if it's the performers messing about then it might have a long name. I have probably called a bunch of things "swingout" when they should be called "sendout" or something. I need someone to check my working.

$count is approximately how far through the move we are (pretending that the move was stretched out to an 8 beat bar). I've found that 7, 8, 1 and 2 of a swingout are really tricky to tell apart, so my plan is to train a model that gets these confused, and then add a layer on top that keeps track of things across time, and use that to improve my dataset.

$direction describes the direction of the follow relative to the lead at the end of the move (if it's a swingout). North is away from the camera, west is to the left of screen, east is to the right of screen, and south is towards the camera.

There are far too many classes for the size of the dataset, I suspect.

Current status

I only tagged a single performance so far, and only 1fps. I need to tag more, then train a model, then do some assited tagging

Links to external resources

Currently taking videos from "All Star Draw" (previously "All Star Jack and Jill") finals, because they are the most natural and fun to watch.

Contribution and labeling guidelines

Please get in touch on facebook or twitter if you want to help.

Supervision

Build Computer Vision Applications Faster with Supervision

Visualize and process your model results with our reusable computer vision tools.

Cite This Project

LICENSE
MIT

If you use this dataset in a research paper, please cite it using the following BibTeX:

                        @misc{
                            dance-moves_dataset,
                            title = { dance moves Dataset },
                            type = { Open Source Dataset },
                            author = { alsuren },
                            howpublished = { \url{ https://universe.roboflow.com/alsuren/dance-moves } },
                            url = { https://universe.roboflow.com/alsuren/dance-moves },
                            journal = { Roboflow Universe },
                            publisher = { Roboflow },
                            year = { 2023 },
                            month = { may },
                            note = { visited on 2024-11-22 },
                            }
                        
                    

Similar Projects

See More