jeremyipark/vision-demos

vision‑demos – Real‑world computer‑vision demos

What it is – A small collection of ready‑to‑run demo applications that showcase how modern pose‑estimation and segmentation models can be applied to everyday activities. Each demo lives in its own sub‑folder with a focused README, installation steps, and a short video‑style illustration.

Demo	What it does	Core model(s)
dance_sync	Takes two videos of the same choreography, overlays the estimated body poses and produces a numeric similarity score that quantifies how well the dancers stay in sync.	vitpose-plus-large
chin_ups	Analyzes a clip of someone doing chin‑ups, counts repetitions, and timestamps the upward and downward phases of each rep.	vitpose-plus-large
rock_climbing	Segments a bouldering wall into individual holds, identifies which holds the climber used and in what order, and compares multiple attempts on the same route.	sam3.1 (segmentation) + vitpose-plus-large (pose)
running	Measures a runner’s cadence, timestamps each foot‑strike, and computes an average knee‑shape profile at contact.	vitpose-plus-large

How it works

• Pose estimation is performed with the vitpose-plus-large model (a Vision‑Transformer‑based human‑pose detector). The model returns 2‑D joint coordinates for every frame, which the demos use to compute timing, similarity, or biomechanical metrics.
• Segmentation for the climbing demo relies on Meta’s sam3.1 (Segment Anything Model) to isolate individual climbing holds on the wall.
• Simple Python scripts stitch the model outputs together, overlay visualisations on the original video, and generate the numeric summaries shown in the thumbnail panels.

Getting started

1. Clone the repository.
2. Follow the per‑demo README in the sub‑directory you’re interested in (e.g., dance_sync/README.md).
3. Install the listed Python dependencies (typically torch, opencv-python, and the VLM model client library).
4. Download the required model weights via the provided links (the README points to the VLM model hub).
5. Run the demo script on your own video file.

Intended audience

• Researchers or hobbyists who want quick, concrete examples of pose‑based analysis for sports, dance, or movement science.
• Developers looking for a template to build custom analytics pipelines around the same models.

License

Apache‑2.0 – free for commercial and non‑commercial use, with attribution.