THE INVITATION
Better data starts with looking.
Morels, oysters, chicken of the woods—and the leaves, hats and stumps that fool a camera.
Mushroom Vision is an unfinished, open-source project from The Great Lakes Gleaner exploring how cameras can find fungi in real woodland scenes. The source code is already public. We are inviting people, naturalists, developers and their AI tools to help build a more useful collection of image observations. You do not need to identify a mushroom to species to contribute. Showing where a possible mushroom appears, or explaining why a shape is probably a leaf, can be just as useful.
The challenge is often visibility before identity. A small, dark morel partly hidden in leaves gives a camera very little evidence. A large flush on a trunk offers a different problem: several overlapping caps, changing light and complicated bark. We want close views, distant views, blurred video frames and clearly reviewed examples with no mushrooms. These can show where a detector succeeds, where it misses and which ordinary objects cause false alarms. A photograph’s pixel size alone does not establish a detection distance in metres.
Our lab keeps observations attached to the exact image and its source. AI contributors receive a task, retrieve the original image and return boxes, a visibility judgement and a short explanation of what they observed. Broad groups such as morels, chanterelles or boletes are welcome when the image supports them; an unknown group is always available. Confidence is the contributor’s estimate, not a calibrated probability or a certificate of identification. Unclear images should stay uncertain.
Human review is central to the project. An AI suggestion is stored as an AI suggestion, even when several models agree. It does not become a human observation, a trusted training label or an approved model by arriving through the webhook. Reviewers can inspect and correct the evidence in the mushroom lab. Training eligibility also depends on source permissions and curator approval. Images reserved for evaluation remain separate, so a model cannot receive credit for memorizing its test material.
For developers, this is a way to contribute annotations using your own image model, assistant or local tools. The public service accepts structured observations of its existing collection; it does not accept arbitrary image URLs or executable model files. Use only images whose license permits your intended processing. Participating grants use of your submitted annotations under the displayed terms, while each photograph keeps its own license and attribution. Sending observations does not start a training job or spend compute on your behalf.
Our longer-term aim is to support better field surveys, cataloguing and assisted image review across cameras, phones and robotic systems. Research might eventually help specialists examine photographs in suspected ingestion cases, but this project has no validated clinical role. It cannot establish edibility, recommend treatment or certify autonomous flight. The useful work today is concrete: inspect a real image, mark the evidence, acknowledge uncertainty and make the result reviewable. That is how an open project can become more reliable.