Preview. Accounts, uploads and points are simulated in your browser while the backend is built. Nothing is stored or sent.
Sheet 01 · Ashimoto Labs · Field Data Program
Every photo you take helps someone walk their city with confidence.
Ashimoto Labs builds Sakimo, a wearable navigation aid for blind and low-vision walkers. It tells the walker what’s ahead by detecting crosswalks, stairs, tactile paving, and obstacles. Every step becomes an informed one. Your photos train the model that makes that information trustworthy.
Real speech, in your browser, right now.
See what contributed data has already done → Impact
Live campaign
First field round: 2,000 in-domain images
1,214 of 2,000 in-domain images collected (demo count). Every accepted photo and every kept walk frame moves the bar. The next training run starts when we hit our target.
Process
How it works
1 · Snap
Photograph crosswalks, stairs and tactile paving as you move through your city or town, during your commute, while running errands, or whenever you have some free time. Take the photos from walking height, the way the device sees them. Faces and license plates are blurred automatically when you upload.
2 · Label
Draw boxes around objects in contributed photos, right in your browser, by mouse or keyboard. Every image is checked by multiple people, so no single mistake gets through.
3 · Train
Accepted images and consensus labels flow into the next model version. Better scores mean fewer missed stairs and surer crossings. You'll see the gains in the monthly newsletter.
Mission
Why this project exists
White canes find what's within reach. Guide dogs are life-changing but scarce. The National Federation of the Blind estimates that only about 5% of people who are blind or visually impaired use guide dogs. Sakimo adds a third layer, a small camera device that speaks the street (“crosswalk, ahead”; “stairs, down, left”), built on the yellow tactile paving system pioneered in Japan and now used worldwide. Our company name comes from the station announcement ashimoto ni gochūi kudasai, “mind your step.” The device is called Sakimo, from saki (先), meaning “what's ahead.” Underfoot, and ahead.
During the development of this project, our biggest hurdle has been training our model on quality data. Street scenes are endlessly varied, and a walker can only trust announcements built on examples the model has truly seen. More real streets in the dataset means more information delivered with confidence. That's where you come in.
Calibration
Detection quality by class, v13
The weakest classes are weak due to insufficient real-world examples.
Missions on the Contribute page target exactly these gaps.
Principle
Built with, not for
Nothing about blind navigation should be designed without blind navigators. What Sakimo announces, and just as importantly when it stays quiet, is shaped by the people who use it. We're forming an advisory circle of blind and low-vision walkers and O&M specialists, and design decisions route through their feedback first.
Join the pilot program on the Support page. Walkers, O&M specialists and sighted allies are all welcome.
Sheet 02 · Field collection
Contribute photos
Upload street photos of crosswalks, road markings, stairs, tactile paving and cones. Accepted photos earn 10 points each.
Non-visual friendly. Photo uploads, Walk mode (with voice and vibration feedback), pilot testing and announcement feedback all work fully without vision. Box labeling is the one sighted task here.
Drop photos here, or browse
JPG/PNG/HEIC · preview build: photos are hashed in your browser and not uploaded yet
No photos handy? Try Walk mode →
Upload queue
Nothing uploaded yet in this demo session.
Field guide
What makes a great photo
- Shoot from chest height, while standing or walking. The device sees the world this way, not from a crouched “photography” angle.
- Include the approach. Capture the same stairs or crossing from 30, 15 and 6 feet away (long, medium, close).
- Odd angles, rain, night, shadows and clutter are more valuable, not less.
- Only upload photos you took yourself. You keep ownership; you grant a training license (contributor terms).
Open missions · July 2026
July missions
The model's weakest spots right now. Mission photos earn a +5 bonus.
Sheet 03 · Annotation bench
Label images
Draw a tight box around each object, pick its class, submit. Points are confirmed once 2 of 3 labelers agree. Quality beats quantity here.
Tutorial. These three practice scenes are drawings, not photos. Learn the tools here risk-free; the real photo queue opens when the site goes live.
Tutorial 1 of 3 · practice scene
Mouse: click–drag to draw. Keyboard: Add box, then ←→↑↓ move · Shift+←→↑↓ resize · Enter place box · Esc cancel. Box position is announced to screen readers as you move it.
Image 1 of 3
Boxes drawn
- No boxes yet.
The label law
Rules from the ontology spec
For any flat or painted ground feature, walk this order:
- Raised domes or bars you'd feel underfoot → tactile paving. Texture, not color. Real pads come in red, yellow and weathered gray.
- Painted bars on the roadway, across traffic → crosswalk.
- Any other paint (lines, arrows, words) → road marking. Painted text is never tactile.
- Tight boxes on the visible part only; one object, one box.
- Do label faint, distant and oblique objects you can confidently identify. Missed faint crosswalks are the number-one field failure.
- Cone = orange traffic channelizer. Post = any rigid vertical, whatever the color. A fluted lamp-post base is a post, never stairs.
- Mind up vs down on stairs; descending is the fall hazard. Genuinely unsure? Skip. A wrong label is worse than none.
Sheet 04 · Results
What the data has done
Every training run gets published here. Contributed photos turn into detection quality, and detection quality turns into safer walks.
Live campaign
First field round: 2,000 in-domain images
1,214 of 2,000 collected (demo count). Photos and walk frames from the community, at deployment scale and angle.
Validation set
Detection quality, v13 (validation mAP 50-95)
Stairs improved 12× between v12 (0.026) and v13 (0.329). That is what focused data does. The next round aims field photos at crosswalks, road markings, tactile paving, cones and descending stairs.
The honest ruler
Validation vs the blind gold set (mAP 50)
In June we built a 315-image blind gold set of real streets and real distances, labeled twice and reconciled. It cut our lab numbers by more than half. To close this domain gap we introduced Walk mode and the 2,000-image campaign. (Posts score ≈ 0 on gold; parked until depth sensors.) With safety being a major tenet of this project, we are pleased to see that v13.1 makes fewer false announcements (precision 0.42 → 0.46).
Coverage · demo
Sheet 05 · Standings
Leaderboard
July 2026 · resets monthly. Points only count once photos pass review and labels reach consensus.
The contributors below are simulated placeholders until accounts go live.
| # | Contributor | Photos | Labels | Mi | Points |
|---|
Scoring
How points work
- +10 per photo accepted in review (auto-blur + human check)
- +5 mission bonus for photos the model needs most
- +5 per label set once 2 of 3 labelers agree
- +25 when a friend you refer gets their first photo accepted
- +1 per dollar donated, credited manually from Ko-fi messages for now
- 0 for rejected, duplicate or off-topic content; quality is the whole game
- Miles are glory, not points, and only count at walking speed (drive, and the meter stops)
Random “gold standard” images with known answers are mixed into the labeling queue to keep everyone honest.
Recruit
Refer a friend
Every new contributor multiplies the dataset. When your friend’s first photo is accepted, you earn +25 points.
ashimotolabs.com/r/guest
Planned
Badges
Cone Hunter · 25 cones Stairmaster · 50 stairs Tenji Friend · 100 tactile
Sheet 06 · Dispatches
Monthly newsletter
One email a month with what your photos and labels did to the model, field-walk stories, and what to hunt for next. No spam, unsubscribe anytime.
Subscribe
Latest issue · sample
Issue #1 · July 2026 · “v13 is live”
- New model deployed on the device: 9 classes, 30 fps, fully on-device.
- Stairs detection improved 12× over the previous version.
- Training moved to on-demand cloud GPUs. Every donated dollar now converts directly into training runs.
- Next up, the first extended field walk and tuning when the device speaks (only what's in your path).
- We're hunting crosswalks, road markings, tactile paving, cones and descending stairs.
Field notes
From the lab notebook
18 Jun 2026
The lamp-post that impersonated a staircase
v13 kept announcing stairs at fluted cast-iron lamp-post bases, at the same confidence levels real stairs occupy, so no threshold can separate them. Looks like we will need more real stairs, and more lamp-post bases labeled as not stairs.
15 Jun 2026
Stairs, twelve times better
One focused data push took stairs detection from 0.026 to 0.329 mAP, a 12× jump in a single model version. Next, descending stairs get their own class, because down is the fall hazard and deserves its own, more urgent announcement.
15 Jun 2026
We retired the copy-paste GPU
Training moved from a free-tier notebook to rented cloud GPUs. A model compile now costs about $0.39 an hour. It means every donated dollar converts to training runs with zero overhead, and we can retrain the moment your photos land.
Sheet 07 · Backing
Support the project
Ashimoto Labs is an independent, open project. There are four ways to help. Photos, labels, testing and, if you like, funding.
01 · Fund
Donate
Donations buy GPU training hours (≈ $2/hr), model-compile pods, image storage, and eventually depth + precision-GPS sensors for the device. You also earn 1 point per dollar.
Donate via Ko-fiOpens Ko-fi in a new tab. Add your contributor name in the Ko-fi message and we credit your points manually until webhooks arrive. Donations are not tax-deductible at this stage; we're exploring fiscal sponsorship so they can be.
02 · Advise
Feedback
Blind or low-vision walker? O&M specialist? Your input shapes what the device announces and how this site works.
03 · Partner
Business & research
Pilot programs, dataset partnerships, municipalities, accessibility orgs, press:
[email protected]
(email forwarding coming online soon)
The contributed dataset is planned for release under an open license, so the whole assistive-tech field benefits, not just this device.
04 · Test
Pilot program
Blind or low-vision walker? O&M specialist? Sighted ally? Help decide what the device announces, and just as importantly, when it stays quiet.
Straight answers
FAQ
Who owns the photos I upload?
You do. You grant Ashimoto Labs a non-exclusive, irrevocable license to use them for model training, including commercially, and the curated dataset is planned for release under CC BY 4.0. The full text lives on the contributor terms page; it is short and written in plain English.
What happens to faces, license plates and my location?
Faces and plates are blurred automatically before anyone sees the image, location metadata (EXIF) is stripped on upload, and every photo passes moderation before it enters the labeling queue.
Where does donation money actually go?
It buys infrastructure. GPU training hours at about $2 an hour, model-compile pods, image storage, and eventually depth and precision-GPS sensors for the device. Training runs are itemized in the newsletter.
Are donations tax-deductible?
Not currently. Ashimoto Labs is an independent project, not a registered charity. We're exploring fiscal sponsorship to change that, and we'll say so plainly here if it happens.
Is Sakimo a medical device?
No. It's an informational aid that complements a white cane or guide dog. It never replaces them, and it makes no medical claims. Think of it as extra information, spoken aloud.
When can I get the device?
It's in active field testing. The honest answer is "not yet." Join the pilot program above and you'll be first to know when that changes.
How do you keep labels accurate?
Every image is labeled by multiple people and only counts at 2-of-3 consensus; gold-standard images with known answers are mixed into the queue; and rejections always come with a reason so the quality bar is learnable.
Sheet 02A · Field capture
Walk mode
Phone at chest height, lens facing the street, and just walk. Walk mode captures frames the way Sakimo sees the world, which is exactly the data the model is missing.
Demo. Frames stay on this page and are thrown away when you leave. Nothing is blurred here because nothing leaves your device. On the live site frames upload with provenance attached, after automatic face and plate blurring at ingest.
Capture
Camera off. Start it, or use Simulate if you're on a desktop.
Form
How to walk
- Chest height, lens level, not aimed at your feet.
- Keep moving. Standing still just makes duplicates, and duplicates are skipped automatically.
- Shoot the boring stuff. Faint paint, weathered pads, and distant crossings are the exact type of frames the model lacks.
- Avoid people where you can; the live pipeline blurs faces and plates anyway.
Session
Frames: 0 · dupes skipped: 0 · 0.00 mi
No frames yet.
Ledger
Provenance, baked in
Every frame is born with its paperwork. Uploaded images are assigned a capture time, session id, contributor, license grant version, a SHA-256 content hash and a 64-bit perceptual hash. Near-duplicates (hash distance ≤ 6) never enter the queue, which protects both the dataset and the points economy.
License reference: Ashimoto Contributor License v0.1.
Document AL-A11Y · Statement · Version 0.1 · July 2026
Accessibility statement
This site exists to serve blind and low-vision users. If any part of it fails you, that is a defect, not an inconvenience, and reports go to the front of the queue.
Standard
We target WCAG 2.2 AA across the whole site, including the interactive labeling bench and Walk mode.
What works today
Everything operates by keyboard, including drawing label boxes (arrow keys move and resize; positions are announced). Page changes, points and status updates are announced to screen readers. The header offers a dark scheme, high contrast (combine the two for yellow-on-black), plain reading mode and four text sizes. System preferences for dark mode and increased contrast are detected automatically, and reduced-motion settings are honored. Walk mode offers voice and vibration feedback so it can be used without watching the screen.
Non-visual ways to contribute
Photo uploads, Walk mode, pilot testing and feedback on announcement wording work fully without vision. Drawing label boxes is the one inherently sighted task; if that ever changes, it will change here first.
Known limitations
This is a prototype. Content is simulated, and it has not yet had an external accessibility audit or full assistive-technology testing. The labeling bench relies on vision. Contrast in the default theme has been checked by design rather than by audit.
Testing commitments
Before launch we commit to NVDA and VoiceOver passes on every page, automated checks (axe) on every release, and testing with blind and low-vision pilot members before public release, not after. Built with, not for.
Report a barrier
Use the feedback form on the Support page or email [email protected]. We aim to respond within 7 days, and accessibility reports jump the queue.
Document AL-CLA · Version 0.1 · July 2026
Contributor License Agreement
The deal, in one breath: your photos stay yours. You give Ashimoto Labs permission to use them to train navigation models, including commercial ones, and to publish them in an open dataset. That permission is permanent for anything already trained or published.
Draft. Version 0.1 has not been reviewed by a lawyer yet and is not legal advice. It will be reviewed before the live site accepts real uploads.
§ 1 · What this covers
Everything you submit: photos, walk-mode frames, bounding-box labels, and feedback ("contributions").
§ 2 · You keep ownership
You keep the copyright in every photo you contribute. This agreement is a license, not a transfer.
§ 3 · The license you grant
You grant Ashimoto Labs a non-exclusive, worldwide, perpetual, irrevocable, royalty-free, sublicensable and transferable license to: store and reproduce your contributions; modify them (crop, blur, augment, annotate); use them to train machine-learning models, including models embedded in commercial products; and distribute them as part of research datasets.
§ 4 · The open dataset
Curated contributions are planned for public release under CC BY 4.0. Attribution is satisfied by a contributors file listing your username (or "anonymous," your choice). Trained model weights are separate works and may be used commercially.
§ 5 · Your promises
You took the photo yourself (or hold the rights to it). You are 16 or older. You will not deliberately photograph identifiable people, private interiors, or anything unlawful.
§ 6 · People and places in photos
Faces and license plates are blurred automatically at ingest, before any human review. Location metadata (EXIF) is stripped on upload. Report anything the blur missed and it will be removed.
§ 7 · Deletion, honestly
You can delete your account and contributions at any time. Deleted contributions leave the labeling queue, all future dataset versions, and all future training runs. What cannot be undone is a model already trained or a dataset version already published. That is why the license is irrevocable for past use; we would rather say that plainly than pretend otherwise.
§ 8 · Labels and points
Labels you draw carry the same license as photos. Points are a thank-you and a scoreboard; they have no cash value.
§ 9 · Changes
This document is versioned. Material changes get a new version number and an announcement in the newsletter, and never apply retroactively to narrow your rights.
Questions? [email protected], or the feedback form on the Support page.
Sheet 00 · Personal
My bench
Your session at a glance. In the live site this persists to your account; in the demo it resets on reload.
Log
Recent activity
Rejections always come with a reason. They teach the quality bar, and they never cost points.
Progress
Session badges
First Steps · one photo accepted
0 / 1
Apprentice Labeler · three label sets
0 / 3
Backer · fund a training run
0 / 1
Account
Devices and recovery
On the live site, your name and points are tied to this device. Link codes carry them to your other devices.
Stored with your account for future recovery. Verified email sign-in arrives in a later phase; until then, link codes are how you add a device.