Veo brought auto-follow filming to baseball. We bring baseball-native computer vision — pitch, ball, and game-state intelligence a repurposed soccer follow-cam can't produce — on a cloud-processed, multi-camera platform that's turn on and go.
The lane isn't "baseball streaming" anymore. Veo Cam 3 now ships a dedicated baseball product with auto-tracking from first pitch to final out. Winning means going deeper than a follow-cam can — into the ball, the pitch, and the game state — where an accessibility gap still leaves most programs unserved.
Market sizing per sports-CV industry reporting (2024–25). An ETH Zurich analysis noted that while elite clubs run data-science departments, most organizations can't implement these technologies — the accessibility gap our price point targets. Veo pricing per public listings, 2026.
Veo and the cheap auto-trackers own broadcast video. Rapsodo, TrackMan and Yakkertech own the analytics. Almost nothing sits where broadcast-quality streaming meets Statcast-style, baseball-native intelligence — that intersection is the opening.
Illustrative positioning. TrackMan, Rapsodo, Yakkertech and Diamond Kinetics lead pitch/swing analytics but not broadcast streaming; XbotGo and BallerCam offer low-cost auto-tracking video with little baseball-specific intelligence, often "no subscription."
Where the two products fundamentally diverge — architecture, ball tracking, hardware economics, and the "turn on and go" experience.
| Dimension | Diamond Vision | Veo Cam 3 |
|---|---|---|
| Ball & pitch tracking | Core design goal — Kalman filter + optical flow + global-shutter 60–120 fps capture built for hard-hit balls and pitch trajectory | Not advertised; tracks the "action" region, heritage is soccer's large, slow ball |
| Area-of-interest tracking | Predictive AOI — a dedicated analysis feed anticipates where the play is heading and drives framing | Single-lens follow-cam pans/crops within one fixed vantage |
| Multi-camera | Synchronized multi-cam with cloud fusion & auto-switching to the best angle per play | Single proprietary dual-lens unit — one vantage point |
| Processing model | Cloud CV (10–30 s latency budget) — no edge compute ceiling, large models, instant updates | On-device follow-cam tracking; tagging runs post-game in the cloud |
| Game-state analytics | Balls / strikes / outs / base state, hit trajectory — Statcast-style structured output | Highlight tagging & clipping in Veo Editor; no game-state engine |
| Setup experience | Pre-provisioned SIM & pre-configured devices — power on, calibrate, connect | Portable single box; 5G model needs a SIM, standard needs Wi-Fi/hotspot |
| Hardware cost | ~$199–250 per camera body; a 2-cam rig undercuts one Veo unit | ~$1,533+ one-time for the camera |
| Software layer | To build — the coaching/editor layer is where stickiness lives | Mature: Veo Editor, drawing tools, clipping, sharing, live streaming |
| Brand & trust | Pre-market challenger | Funded incumbent with pro-club logos |
The hardest CV problem in the sport — a 95+ mph liner crossing the frame in a few hundred milliseconds — is our starting point, not an afterthought. A small, fast ball at high velocity is exactly what a soccer-derived follow-cam can't retrofit.
Synchronized, keyframe-aligned cameras give field-wide coverage and the geometry to convert pixels to real-world measurement — then auto-cut to the angle that best shows each play, like an automated broadcast truck.
Commodity C-mount global-shutter bodies at $199–250 mean a full multi-cam rig costs less than one Veo unit — decisive in price-sensitive youth and travel-ball programs that the accessibility gap leaves unserved.
The 10–30 second streaming-latency budget is the decision that unlocks everything downstream. It moves CV off the camera and into the cloud, which is what lets cheap hardware punch far above its price — and what a single fixed-function device structurally can't match.
Global-shutter cameras, 60–120 fps, distortion-free on fast motion. One analysis feed drives framing; others stream.
H.265/AV1, CV-optimized. ROI encoding keeps the field sharp, compresses the crowd. Idle-state bitrate drop on dead ball.
SRT ingest → FFmpeg → HLS 4–6 s segments to object storage, plus CV frames to the queue. Keyframe-aligned via NTP.
Stateless pods: detection, ball tracking, pose — each an independent service, autoscaled on queue depth, scale-to-zero between games.
ABR playback to viewers; game-state and trajectory to the analytics store, feeding highlights and the coaching layer.
Full pipeline: Camera → SRT (bonded cellular) → MediaMTX ingest → FFmpeg/NVENC transcode → HLS to object storage + CV frames to Kafka → GPU inference workers → results stream → analytics, overlays, storage.
Because streaming tolerates 10–30 s latency, inference runs in the cloud, not on a Jetson at the field. That removes the edge compute ceiling — enabling large models, multi-camera fusion, and instant model updates with no truck roll. Veo's on-device tracking is capped by exactly the hardware limit we design around.
Bonded cellular + Wi-Fi gives a variable 4–8 Mbps per camera, so we never stream raw. H.265/AV1, ROI encoding (sharp field, compressed crowd), and idle-state bitrate reduction on dead ball keep multi-camera streaming feasible where a single 4K feed would not be.
Every HLS segment and CV unit must be independently decodable — requiring NTP clock sync, forced keyframe intervals, and identical segment durations across all cameras. Get it wrong and you get stalls and false detections at boundaries. It's unglamorous, hard to replicate, and the backbone of reliable multi-cam.
Kubernetes/Helm, Kafka, MinIO (S3 API), Triton, KEDA, Terraform — the same charts run on minikube or any cloud. No managed-service lock-in means we deploy to the cheapest compute and preserve gross margin as we scale, rather than renting a hyperscaler's proprietary pipeline.
The one dependency to manage: YOLO-family detectors ship under AGPL-3.0, which forces open-sourcing or a commercial license for closed products. Apache-2.0 alternatives (RF-DETR, YOLO-NAS, YOLOX) sidestep it — a licensing choice that shapes the productization path below.
The goal isn't more capability at the cost of friction. It's more intelligence with less setup than the incumbent — the multi-camera analysis feed feeds automated direction, not manual work for a volunteer parent.
A dedicated analysis camera reads pitch, contact, and fielder movement to point the frame where the ball is going, not where it just was — the master/slave switching pattern, driven by CV.
Cloud fusion of synchronized angles, auto-directed. The system cuts to whichever camera owns the play — infield, baseline, or outfield — like an automated broadcast truck.
Cameras ship with SIMs installed and devices configured. No hotspot pairing, no network setup. Power on, and the rig calibrates, bonds its modems, and starts.
Every play resolves to structured data — count, outs, base state, trajectory — written to the time-series store, feeding highlights, analytics, and coaching automatically.
Beyond Veo, two tiers matter: cheap no-subscription trackers attacking the low end, and baseball-analytics incumbents who own the data but not the video.
The clearest lesson from CV companies that scaled: permissive licensing maximizes adoption but minimizes capture; fully restrictive licensing does the reverse. The sweet spot — Roboflow's playbook — is open-core.
Release libraries and base models under Apache-2.0 to drive adoption and community goodwill. Network effects and switching costs compound as teams build on the ecosystem.
Models are free; the annotation, calibration, managed inference, streaming, and coaching/editor layer are paid. Sell the platform, not the model — that's where stickiness and recurring revenue live.
License the CV models under AGPL-3.0: anyone can use them, but companies building commercial products must open-source or buy a license. It captures enterprise value without becoming free R&D for larger competitors.
"It's not an open-source problem, it's a business problem." Free models + paid platform (or permissive tools + AGPL models) lets a CV company grow a community without handing its work to better-funded rivals.
Our edge is not baseball streaming — Veo covers that. Our edge is baseball-native computer vision: predictive tracking, intelligent multi-cam switching, and game-state analytics that a repurposed soccer follow-cam can't produce — enabled by a cloud-processed, vendor-neutral architecture running on cheap, pre-provisioned, turn-on-and-go hardware.
The coaching/editor layer is where stickiness lives — and it's the whole thing we still have to build.
Veo set the "one box, minutes to set up" expectation. A multi-cam bonded-cellular rig must feel just as effortless.
Sub-$500 no-subscription trackers can eat the bottom of the market before we reach it.