Dancing With Myself
summary
Concrete claim: models trained on personal motion capture actually generate realistic, style-specific choreography and useful variations. Tools and data: a VAE and a graph-based GNN operate on a 20-camera capture with about 50 motion points at 35 fps and a 256-d latent space. Built artifacts: fully generated sequences, a recorded duet with the same performer, live-feedback experiments and an edge-analysis that finds surprising cross-body links (right hand to left foot). Recommendation: worth watching for methods, results, and concrete failures.
AI-written summary · about
chapters (14)
- 0:00Start
- 3:15Why use generative models in dance?
- 4:00Previous Work
- 5:06Input data samples
- 7:12VAE for Fixed-Length Movement Sequences
- 8:23Sequence VAE: Random sampling
- 11:29Variations on Existing Data
- 16:03Graph Neural Network Study Motivation
- 16:55Graph Neural Networks
- 17:22NRI: Neural Relational Inference
- 18:10Model Architecture
- 19:11Edge Type Visualizations
- 22:58How can I use these models?
- 24:22Glitches are so appealing!
related talks
39mChoreorobotics: Teaching Robots How To Dance
29mOf Muppets, Metaballs, and Ballroom Dance!
40mOf Mouse and Man: Creative Networks and Mechanical Directors
38mMaking Art with Artificial Intelligence: Artists in Conversation
06mCoolMoves: User Motion Accentuation in Virtual Reality
28mVibing Vision: how we built a pose tracking art installation in two weeks
45mGenerative AI for Image and Video Production
10mDeepNAG: Deep Non-Adversarial Gesture Generation