Infrastructure
Inference & Deployment
The deployment design: video → detection → pose → model in a single forward pass, with quantization and distillation goals for edge-class hardware.
The pipeline, as designed
The deployment design keeps the path from camera to insight short. Video is decoded and players detected and tracked; a pose estimator turns each player into per-frame keypoints; windows of those keypoints enter the foundation model, and one forward pass returns both outputs — a clean 3D skeleton and the movement embedding. No stage in the design requires special hardware on the pitch: the input is ordinary video.
Small enough to move
The backbone is a deliberately small model by foundation-model standards (see the architecture datasheet), and the deployment plan pushes further: post-training quantization to shrink memory and latency, and distillation into task-specific student models where a product needs only a slice of the model’s understanding. The goals are stated as goals — measured latency and throughput numbers will be published when a serving pipeline exists to measure them.
Design goals
| Goal | Rationale |
|---|---|
| single-pass serving | Both outputs of the model from one forward pass — no chain of specialist models to operate. |
| quantization | Post-training quantization to cut memory and latency on commodity GPUs. |
| distillation | Task-specific student models where a product needs a fraction of the backbone. |
| edge-class targets | Analysis near the camera where bandwidth or privacy requires it. |
| batched pose front-end | Amortized per-player cost when many bodies share one frame. |
The query surface this pipeline is designed to feed is specified in the knowledgebase API.