Separate the vision jobs
Scene segmentation supplies runway and pit geography; the vault-phase model remains responsible for finding the active vaulter and understanding the attempt.
Computer vision + live operations
PV Video Capture
A multi-camera capture and clip-delivery system with two custom computer-vision models and more than 30,000 labeling-ready frames.
StatusLive private operational system
PlatformsDesktop, web, mobile
RoleProduct owner and software engineer
SourcePrivate repository
DemoIsolated dummy athlete account
The product
Built around a real workflow.
This system connects physical cameras and a live runway workflow to software that identifies attempts, manages athletes, creates clips, supports model-assisted labeling, and delivers video. I built the data pipeline and trained separate models for scene segmentation and vault-phase detection, combining local, latency-sensitive Python tooling with a hosted FastAPI application and cloud media storage.
- Multi-camera capture, PTZ controls, live previews, attempt monitoring, automated clipping, and recovery-aware launch tooling.
- A runway/pit segmentation model and a multi-class vault-phase model trained with versioned datasets, same-split evaluation, and controlled checkpoint promotion.
- Athlete accounts, practice check-in, clip assignment and playback, mobile access, and self-hosted primary/secondary server failover.
Model training
Two trained models. One field-ready vision pipeline.
The scene-segmentation model learns the physical geography of the runway and landing pit. The vault-phase model detects the active athlete and classifies phases such as the run, plant, takeoff, top of jump, landing, and walk-off. Keeping those responsibilities separate makes the system easier to evaluate and safer to improve.
The live labeling workspace now contains 30,605 training-ready frames: 22,027 positives and 8,578 negatives. The current published vault-phase model was trained and validated on a 15,579-image split, while the segmentation model was refined on 397 carefully reviewed polygon masks.
30,605labeling-ready frames in the live workspace
15,579images in the current vault-phase split
0.947vault-phase validation mAP50
0.952scene-segmentation mask mAP50
- 01
Extract candidate frames from real practice and meet video.
- 02
Label and review phase boxes, negatives, runway masks, and pit masks.
- 03
Train and validate YOLO checkpoints against stable dataset splits.
- 04
Publish a candidate only when it beats the current same-split baseline.
Architecture
From user action to production service.
A public, high-level view of the system. Credentials, private data, deployment identifiers, and source code remain protected.
01Cameras + runway
PTZ and fixed cameras capture live attempts
→02Local vision pipeline
Python, FFmpeg, OpenCV, segmentation, vault phases, Face ID, and clipping
→03Self-hosted FastAPI
Accounts, check-in, roster, labeling APIs, and two-server failover
→04Shared data + media
Neon/Postgres, Cloudflare Stream, and R2
→05Athlete + staff tools
Browser, Expo mobile, check-in station, and labeling workspace
Technology stack
PythonOpenCVUltralyticsMediaPipeFFmpegFastAPIExpoCloudflare StreamR2Neon
Engineering decisions
What makes the system hold together.
Measured model promotion
Candidates are evaluated on the same split as the published checkpoint and promoted only when validation mAP50-95 improves.
Local where latency matters
Capture and inference stay near the cameras, while account, roster, model-labeling, and clip metadata are coordinated by hosted services.
In the field
The physical environment is part of the system.
Camera position, lighting, athlete motion, hardware recovery, and capture timing all shape the software design.
