Study 01 under way · No results yet
PortMind collects frames from public port cameras, keeps every original untouched and turns them into frozen datasets that people label with AI help. The aim is detectors you can trust on a camera they have never seen, with a record of how every result was made.
Illustration · not real camera footage or study evidence
How it works
One pipeline from camera frame to tested model.
Collection never stops. Studies work on frozen copies, so tomorrow's captures cannot change yesterday's result. Every study runs through the same five steps.
Collecting from public port camera feeds in Montréal, Vancouver, Houston and Baltimore.
Capture
Originals are saved before any model looks.
A collector polls public port camera feeds on a schedule and stores the original image bytes first. A saved file is never overwritten. If a camera sends the same frame again, it is skipped.
Catalog
Every frame can be found by camera and time.
Each new file adds a row to a catalog: which camera, when it was captured and where its bytes live. A reconciliation job fills in rows that were missed.
Freeze
Studies run on snapshots that cannot change.
A study picks cameras and dates, then locks that selection with each file's SHA-256 hash. Collection keeps going, so new captures never change an old result. Test frames are reserved before anyone looks at them.
Review
AI suggests. A person decides.
Reviewers label frames in a private app. Moondream proposes boxes and Llama 4 Scout gives coaching notes. Suggestions are drafts, never truth. The study plan gives each test frame two independent human reviews with AI help turned off.
Evaluate
Tested on cameras it has never seen.
Models train on some cameras and are tested on frozen frames from cameras held out of training. Each run records its code, settings, data and labels. A failed run stays on the record. A new attempt is a new run.
Rules every study follows
Originals first
Source images are stored before inference, so every label can be traced to untouched bytes.
Suggestions are not labels
Model output is a draft until a person accepts or fixes it.
Tests are set before results
The question, the test frames and the metrics are written down first.
Negative results count
Runs where a method did not help are reported too.
Study 01 · Now
How few labels does a new camera need?
A truck detector learns from two port cameras and is tested on two other cameras at the same port. We compare two ways to spend the same small label budget: frames from the new cameras, or more frames from the cameras it already knows. Our guess is that new-camera frames help more. If the guess is wrong, we will say so.
- 680 development frames frozen
- 32 frames in a first calibration pass
- 0 labels approved, no results yet
Option A
More frames from the training cameras
10 extra labeled frames from cameras the detector already knows
Option B
Frames from the new cameras
10 labeled frames from the cameras it is tested on
Same test for both
Both options get exactly 10 extra frames. Then both are tested on the same frozen frames from the new cameras, from a week chosen in advance.
Gain over the base detectorNot measured yet
Planned budgets: 10, 50 and 100 frames. Final budgets follow the pilot.
It can show
Whether a few labels from a new camera help a truck detector more than the same number from cameras it already knows. One port, two new cameras, one kind of object.
It cannot show
How a detector does at a different port, or whether a product could count trucks reliably. Those need a larger, separate study.
What we will publish when results exist
Results per camera
Missed trucks, false alarms and count error on each new camera, shown separately. Results where labels did not help are reported too.
Labeling effort
How long the labels took, not just how many there were.
The data record
Camera IDs, capture times, file hashes, frames left out and why, and every human review decision.
Where it can go
Same pipeline, harder questions.
These are plans, not features. Each one depends on rights, labels and results we do not have yet.
Transfer across ports
Test whether a detector trained at one port works at another. The archive already collects from several ports. A fair study needs usage rights and human labels first.
More studies on one archive
Camera drift, capture quality, duplicates and weather can each be a study on the same frozen snapshots.
Benchmark your model
Let a team run its own model on a frozen, held-out test set and get the full run record back.
Trained reviewers
A reviewer ladder from calibration to domain expert, so hard frames go to people who can judge them.
Forecasting, then simulation
Use camera-derived counts to forecast near-term port activity. Simulation and reinforcement learning come only after forecasts hold up.