Golf swing video datasets: coaching, club tracking, event timing and review

Define the swing-analysis task before requesting footage

Golf video can support swing-event timing, body pose, club tracking, ball detection, coaching retrieval or audiovisual explanation. These tasks need different labels. A clip that shows a player clearly may not resolve the clubhead at impact. A slow-motion export may improve visual inspection but alter the relationship between video time and real movement. Specify whether the model should locate events, track image-space geometry, compare instructional examples or predict a separately measured outcome.

Capture viewpoint is central. Face-on, down-the-line, oblique and elevated views reveal different motion and occlusion. A front-facing clip cannot be assumed to provide the same geometric evidence as a calibrated side view. Define acceptable framing, distance, camera height, frame rate and visibility of the player, club and ball. Preserve complete swings with setup and follow-through where needed rather than selecting only the visually easiest portion.

GolfDB and current public access needs

The original GolfDB paper introduces 1,400 swing videos with event frames and other contextual labels. It is a useful reference for sequencing and metadata design. It does not establish that its source videos carry the same permissions or measurements a commercial coaching product requires. Review the actual release and source conditions before treating it as a purchasable collection.

A September 3, 2026 public GitHub request describes a golf-vision pipeline seeking more robust club-pose data, particularly outside its strongest impact window. That is concrete evidence of a technical need stated by one requester, not a verified sale, budget or Fiund relationship. It suggests that backswing and follow-through coverage can matter as much as a single visually dramatic impact frame. The requester's claimed performance and hardware are not independently validated here.

Potential buyers and source routes

Potential users include coaching applications, swing-analysis developers, sports-vision researchers and training-content search products. Their requirements should not be conflated with general sports highlights. A sequencing model may benefit from many independent players and viewpoints; a club-tracking product may need high temporal detail and difficult occlusions; a coaching assistant may need explanations tied to exactly the frames that support them. These are use-case segments, not named Fiund customers.

Potential sources include original coaches, authorized player recordings, golf-content producers, training facilities and commissioned capture. Source control and participant permissions need review for the intended use. A coach can possess footage without necessarily holding every downstream right a model project needs. Likewise, a public video’s availability does not establish permission for commercial training or a license to the player's likeness. Keep technical source suitability and the proposed permission scope as separate questions.

Temporal detail and capture quality

Frame rate is meaningful only alongside shutter behavior, compression, target size and export history. A high-frame-rate file can still contain duplicated frames or motion blur that hides the clubhead. If a video was slowed, interpolated or re-timed, record the transformation. The labels should specify whether event indices refer to the original capture, a standardized derivative or playback time. For precise timing, inspect the actual frame sequence and timebase rather than trusting a filename.

Framing should preserve the parts needed for the task. A player-centered crop may remove the top of the club or follow-through; an impact-focused crop may lose body posture. Background contrast can make a ball visible in one setting and nearly invisible in another. Include indoor simulators, outdoor ranges, different clothing and lighting only where relevant to the intended deployment. Describe the source distribution honestly rather than treating a few diverse sample clips as proof of broad coverage.

Event definitions, geometry and a record schema

Define the event taxonomy through visible cues. Setup, backswing transitions, top, downswing, impact and follow-through can be labeled under different conventions. Use reviewed examples and permit uncertainty where a cue is not visible. For keypoints, specify the object landmark, coordinate frame and treatment of occlusion. A club shaft can flex and blur; a single image-space line should not be described as a physical measurement without a justified method.

For outcome-linked work, preserve the pairing process. A launch-monitor row is not reliable ground truth if it cannot be connected to the correct swing and time. Similarly, a coach’s qualitative comment should be labeled as interpretation rather than a measured joint angle. Keep raw observations, derived measurements and coaching conclusions separate. This allows a buyer to test which supervision is justified instead of inheriting a stronger claim from a convenient spreadsheet.

Pricing and a worked planning model

No representative public transaction sample establishes an average golf-video AI-training license rate. Costs can include original rights, player recruitment, facility access, camera setup, synchronization, specialist annotations, measurement linkage and review. Existing coaching archives and newly commissioned calibrated capture are different products. Quotes should state whether the unit is a swing, accepted clip, player-session, camera-hour or complete project.

For a hypothetical annotation budget, consider 1,200 swings with eight event anchors each. If initial event review takes an assumed two minutes per swing and independent review takes another minute, that is sixty hours. At a chosen loaded rate of $35 per hour, that component is $2,100. It excludes footage rights, collection, club/ball geometry, expert coaching interpretation, external measurements, corrections and delivery. These are explicit scenario assumptions, not observed rates or a Fiund quote.

Dense tracking can cost very differently from sparse event anchors. A proposal that includes a few keyframes should not be compared directly with one containing per-frame club observations. Ask how uncertain or occluded frames are represented, what fraction is independently reviewed and how many unique players and sessions support the total. A large number of repeated swings by one player may suit a narrow experiment while failing a generalization requirement. Compare the useful coverage, not only the count.

Comparing archives and new collection

An existing coaching archive can contain authentic explanations and natural variation, but may have inconsistent views, clipping, timing and measurement links. New capture can follow a defined camera and metadata plan but needs recruitment, access and a realistic assessment of capture yield. Neither should be represented as automatically superior. The choice depends on whether the project values natural coaching context, precise geometry, broad subject variation or a specific sensor relationship.

A small pilot should test the most expensive assumptions first. If the task requires visible clubhead detail across the whole swing, inspect difficult backswing and follow-through frames before labeling a large batch. If external measurements are essential, verify pairing on a complete session. If coaching explanations are central, check whether the speaker identifies visible evidence or relies on context outside the clip. Use those findings to refine scope and quote units before scaling.

Sample acceptance and evaluation design

Sample by viewpoint, club type where supplied, player group, camera, lighting, speed and source quality. Include occluded clubs, small balls, cropped follow-throughs and ambiguous impact frames. Review originals and the exact model-input derivatives. A resized clip may be adequate for body pose while erasing the clubhead. File integrity, temporal fidelity, object visibility, labels and source documentation should have separate acceptance criteria.

Keep player and session groups together according to the intended test. A random split of adjacent swings can reward a model for recognizing the same setting or player rather than general swing structure. If the product must handle unfamiliar cameras, include that holdout explicitly. Distinguish within-player improvement tracking from new-player generalization. State limitations when a source cannot provide reliable subject grouping rather than treating missing grouping as independence.

A worked frame-review example

Imagine a clip with a clear body silhouette but a clubhead that disappears against a bright background near the top of the swing. It may pass a body-pose brief while failing full-swing club tracking. The acceptance record should name that difference instead of calling the entire clip good or bad. If the buyer only needs event timing, the uncertainty at a particular anchor may be acceptable under a defined rule; if it needs dense geometry, it may require another view.

Compare original timing with a slowed social export and retain their relationship. Check whether labels were placed on actual captured frames or on interpolated playback frames. If a coaching comment refers to an external measurement, verify the exact swing link rather than assuming session proximity proves the pairing. This kind of pilot can reveal which extra source evidence is worth paying for before dense annotation begins.

Custom collection is a separate product from a benchmark

A buyer can also discuss a commissioned-capture route with a provider such as LXT’s video collection service. Its general offering does not establish golf-specific inventory, external measurements or access to any player or facility. A project-specific proposal would need to identify the capture plan, participants, source permissions, timing, labels and pilot criteria. Compare that scope with an archive license and with a public benchmark under its own terms instead of treating them as equivalent purchases.

Deals and a Fiund golf-data brief

No verified golf-video-specific AI-training license deal amount is established by the reviewed sources. Coaching subscriptions, simulator equipment prices and sports media contracts are not reliable dataset price comparables. The public access request described above establishes a technical interest only; it does not validate an offered license or a buying budget.

Send Fiund the model task, required viewpoints, club/ball detail, event or tracking labels, independent player/session targets, external measurement needs and intended uses. Include examples of acceptable and unusable framing. Fiund can assess potential source fit, permissions and the technical review needed for a representative pilot, then discuss a project-specific proposal. Availability, source documentation, measurement links, preparation, acceptance and pricing are confirmed for selected material rather than promised by this category page.

Questions buyers ask