RaceLabs · Performance

Two walls.
We skip both.

Every photo sorter is gated by one of two things: the machine in front of you, or your upload. Local apps can only go as fast as your laptop. Cloud apps make you push every full-size RAW up the wire. RaceLabs is built to dodge both: you upload about 1% of the bytes, your originals never leave your disk to be sorted, and the heavy work runs on datacenter GPUs. A full trackday lands in minutes.

METHOD: COMPRESS HERE, PROCESS IN THE CLOUD UPLOAD: ~1% OF ORIGINAL BYTES ORIGINALS: STAY ON YOUR DISK
~1%
of the bytes
ever leave your card
300/s
JPEGs compressed
on your machine
7,000/min
sorted on
datacenter GPUs
0
full-res originals
sent to analyse
In one line

A local sorter is capped by your hardware; a cloud sorter that uploads your files is capped by your bandwidth. RaceLabs builds a lightweight preview of each photo, roughly 1% of the original size, and sends only that to datacenter GPUs, while your full-resolution originals stay on your disk. So your internet stops being the bottleneck, the heavy identity work never touches your laptop, and a card that would take hours to upload as RAWs is on its way in minutes.

01 · The bottleneck

Everyone hits a wall.
It's just a different one.

"Thousands of photos in minutes" sounds the same on every website. What it never says is where the time actually goes, and that's decided by architecture, not marketing. There are only three ways to build this, and two of them put the wall on you.

  Local desktop toolsrun on your machine Cloud tools that upload originalsprocess on their servers RaceLabscompress here, process in the cloud
Heavy work runs on your laptop / PC their servers datacenter GPUs
You're gated by your CPU / GPU your upload bandwidth neither
Bytes you upload none 100%, every full RAW ~1%, a small preview
Originals leave your disk to sort no yes no
Scales past your own machine no yes yes

A local tool can never beat the silicon in your bag, and the paddock is not where your workstation is. An upload-everything cloud tool can never beat your venue wifi. RaceLabs is the only column that doesn't hand you the wall.

02 · The upload, to scale

You send 1%,
not 100%.

Each ~50 MB original becomes a compact preview of roughly a quarter of a megabyte before anything is sent, about 1% of the size, and the entire difference between a half-day upload and a coffee break. Building those previews flies from JPEG or RAW+JPEG (~300/s); RAW-only is slower to develop, which we're straight about below. Here's the upload, both ways, on the same connection.

Upload all 10,000 RAWs — ~500 GB (other cloud tools)
≈ 5.5 hours
Upload 10,000 previews — ~2.5 GB (RaceLabs, ~1%)
under 2 minutes

Same 10,000-photo day, same 200 Mbps connection. The RaceLabs bar really is that thin, it's not drawn to scale because at scale you wouldn't see it. And the gap is connection-proof: halve your speed and both bars simply double, uploading 1% of the bytes is ~100× faster than uploading all of them on any line. After upload, the sort runs on datacenter GPUs, not the laptop in your bag.

The whole card, timed

Compress — 10,000 JPEGs (or RAW+JPEG) at ≈300/s: about 30 seconds.
Upload — ~2.5 GB of previews on a 200 Mbps connection: about 100 seconds.
Sort — on datacenter GPUs at ≈7,000/min: about 1.5 minutes.
Under four minutes, shutter to gallery — while the upload-everything route is still hours from finishing.
RAW-only? Developing RAWs is slow by design (~5/s), so local prep on a 10k-card is nearer 30 minutes, shoot RAW+JPEG to keep the fast path. Upload and GPU sort are identical either way.

Video keeps the pace

RaceLabs doesn't stop at stills. Footage is processed fast enough that per-rider video clips land alongside the photos instead of holding up delivery — a session with video isn't a second, slower job, it's part of the same ten-minute turn.

03 · What it buys you

Sell while
the adrenaline's hot.

Shutter-to-gallery, not photos-per-minute

The number that pays you isn't engine throughput, it's how fast a session becomes a gallery you can sell. Tiny uploads plus datacenter GPUs collapse that window to minutes, so riders can buy before they've left the paddock, while they still have the lap in their veins.

No paddock workstation required

The identity work, the genuinely heavy part, runs in the datacenter, not on your machine, so a modest laptop handles the sort and you're not lugging a tower to the circuit. (Local preview prep is light for JPEG; RAW-only adds develop time, see the timeline.)

Your originals stay yours

Only the lightweight preview is sent for the sort, full-resolution RAWs never leave your disk to be analysed. The cloud sends back a sort map, which is applied to your local files. Full-size images upload only if you choose to publish a public gallery, never for the sort itself.

Scales with the event, not your hardware

A 2,000-photo club round and a 20,000-photo national weekend take the same shape: small upload, GPU sort, results back. The work scales on our side, so a bigger event doesn't mean a slower laptop or a longer night.

04 · The real clock

What "minutes"
doesn't hide.

Two things worth putting on the table, because a skeptic will ask, and they're both fine once you understand them.

The human pass is a merge, not a bin

A skeptic will say "four minutes hides a manual pile." Two honest things: genuinely ambiguous frames go to a small Unclear bin for a two-second glance, and when our published recall dips, that's not photos in a reject pile, it's a rider occasionally split across two folders (head-on vs from behind) that merge in one click, often automatically by race number. Precision stays 99.6%, so the leftover work is merging, not re-tagging. Full numbers on the research page.

Compression doesn't cost accuracy

A fair question: if you only upload 1%, does the sort suffer? No, identity is read from appearance, which survives downscaling, and our published mAP of 0.99+ is measured on this very pipeline, not on untouched RAWs. You get the small upload and the accuracy, not a trade between them.

The fast part was never the AI. It's that you never had to upload the heavy files in the first place.

The whole idea, in one sentence
05 · Questions

Speed, answered.

Won't uploading thousands of RAW files be painfully slow?
You don't upload the RAWs. RaceLabs builds a compact preview of each photo, roughly 1% of the original size, and uploads only that. A 10,000-photo day that would take around 5.5 hours to push as ~500 GB of RAWs over a 200 Mbps connection uploads as ~2.5 GB of previews in about 100 seconds, and because it's 1% of the data, that gap holds on any connection. Your upload stops being the bottleneck.
Is the speed limited by how powerful my laptop is?
No. The heavy work runs on datacenter GPUs, not your machine, so a modest laptop on track wifi is enough. This is the opposite of a local desktop tool, where throughput is capped by whatever silicon you brought to the paddock and can't be moved off your machine.
What if I shoot RAW, does that slow it down?
Yes, and we're straight about it. From JPEG or RAW+JPEG, preparing previews flies, roughly 300 a second, about 30 seconds for a 10,000-photo card. RAW-only is slower by design: developing full RAWs runs around 5 a second, so that local step on a 10,000-RAW card is closer to half an hour. The fix is simple, shoot RAW+JPEG and you keep the fast path. Either way, the heavy work, identifying every rider, runs on datacenter GPUs and never touches your machine.
Do my originals leave my computer?
Not to be sorted. Only the lightweight preview is sent for analysis; your full-resolution originals stay on your disk, and the cloud returns just a sort map that's applied to those local files. Full-size images are uploaded only if you choose to publish a public gallery, never for the sort itself.
How is this faster than a local tool like RaceTagger?
A local tool keeps everything on your machine, which is great for privacy but means it can only run as fast as that machine, and it can't borrow more power for a big event. RaceLabs runs the heavy work on datacenter GPUs and keeps uploads tiny, so it scales past your hardware without making you push full files up the wire. Different trade: they pin everything to your laptop; we pin almost nothing to your connection.
Does compressing the upload hurt sorting accuracy?
No. Riders are identified from appearance, helmet livery, kit, machine, which survives downscaling, and our published accuracy (mAP 0.99+, 99.6% same-rider precision) is measured on this same analysis pipeline. See the research page for the full validation. You get the small upload and the accuracy together.
What's the realistic shutter-to-gallery time?
For a 10,000-photo trackday shooting JPEG or RAW+JPEG: roughly 30 seconds to compress on your own machine (≈300/s), about 100 seconds to upload the previews on a 200 Mbps connection, and ~1.5 minutes to sort on datacenter GPUs (≈7,000/min), under four minutes shutter-to-gallery, plus a short human pass over the small Unclear pile. RAW-only adds local develop time (~5/s), so shoot RAW+JPEG for same-day speed, the upload and the GPU sort never depend on your hardware.

Time it on
your own card.

Bring a full day. Time it end to end, shutter to a gallery you'd hand a client. That's the only benchmark that's about your shooting, and it's the one we want you to run.