Computer vision platform
Upload a few sample photos and describe what you need. VisionDock suggests the task and labels, checks your data, configures and runs training, and serves the model over a REST API. No ML team, no servers to manage.
Real screen from VisionDock, not a mockup.
Why VisionDock
Picking a task type, defining a label set and building a pipeline usually needs a computer vision specialist.
With VisionDock: upload 10 sample photos and a short chat suggests the task type and labels. One click turns the conversation into a project plan.
Folder layouts, CSV headers, YOLO/COCO/VOC annotations… one mistake wastes an entire job.
With VisionDock: every task type is checked before training. Missing groups, too few photos and broken files are flagged up front.
Compute, endpoints, API keys, scaling, cost tracking — more work than the model itself.
With VisionDock: training runs on managed cloud compute, and the model becomes a REST endpoint in one click — without standing up your own infrastructure.
How it works
Each step builds on the previous one. The screens below are from a real project, a greenhouse tomato leaf classifier, built in the product.
Upload 10 sample photos and write your goal in your own words. A vision-language model looks at the photos, asks a few clarifying questions and turns the conversation into a project plan.
The expected format is spelled out for your task. As you upload, VisionDock checks that every group has enough photos and tells you exactly what's missing. Prefer ready-made data? Import a library dataset in one click.
healthy/, septoria/)
Before training, a vision model inspects your dataset and picks the settings: image size, augmentation, preprocessing and confidence threshold. Review or change anything, then start the run and watch its progress.
Push the trained model to a live endpoint in one click. Test it inside the app, create an API key and connect it to your own systems. Working in the field? Download the edge bundle.
$ curl -X POST "$ENDPOINT/predict" \
-H "X-API-Key: $VISIONDOCK_API_KEY" \
-F "image=@row12_leaf.jpg"
{
"label": "septoria leaf spot",
"confidence": 0.94,
"scores": { "healthy": 0.02, "septoria leaf spot": 0.94, "bacterial speck": 0.04 }
}A real project, start to finish
We ran VisionDock on real photos of greenhouse tomato leaves. Everything below is what the product actually produced. Nothing was staged.
10 sample photos, 2 messages. The assistant spotted “small brown spots and discoloration”, recommended image classification and asked about labels, camera setup and speed.
Generated in one click: 3 classes (healthy, septoria leaf spot, bacterial speck), folder-per-class format, EfficientNet-B0 suggested.
18 labelled photos uploaded into 3 groups. Each group passed the 5-photo minimum before training was allowed.
Chosen by AI from the images: 224×224 input, 40 epochs, batch 32, RandAugment + Mixup + CutMix, denoise and contrast boost.
Sample photos: Wikimedia Commons, released under CC0 / public domain. Production models need far more than 18 photos; this run shows the workflow, not final accuracy.
Find scratches, dents and missing holes, with their location, from a fixed camera on the line and pull defective parts immediately.
Predict a 0–100 ripeness score per fruit, flag anything below the threshold and plan the harvest with the edge bundle running offline.
Count parcels on the belt, flag crushed or opened ones and notify your warehouse system over the API. SAHI helps with small objects on wide cameras.
Task types
Discovery recommends the right type for you. All you have to bring is the right data.
One group per photo.
Examples: Healthy or diseased leaf. Part “OK” or “Defect”. Document type.
Data: one folder per group
Several tags per photo.
Examples: “Scratch + rust + stain” in the same image. Scene attributes. Clothing attributes.
Data: photos + label CSV
Predict a number from a photo.
Examples: Fruit ripeness score. Coating thickness. Fill level.
Data: photos + numeric CSV
Where a single object is.
Examples: Label position. License plate location. Fiducial marker position.
Data: box-annotated ZIP (YOLO/COCO/VOC)
Multiple objects: location + class.
Examples: Every defect on the line. Shelf product counts. Parcels on a conveyor.
Data: box-annotated ZIP (YOLO/COCO/VOC)
Upload a handful of sample photos. The discovery chat recommends a task type and explains why — before any training starts.
Libraries
Ready-to-use datasets organized by industry and task type. Pick one and VisionDock creates a project with the data loaded and ready for training. No upload step needed.
Proven architectures such as EfficientNet and YOLOv8, each with recommended settings for a task. Start a project from a card, name your own groups and train on your data.
FAQ
No. The discovery chat recommends the task type and labels, and training settings are chosen automatically from your images. Advanced users can still change every parameter from the same screens.
VisionDock requires at least 5 photos per group to start, but that is a floor, not a target. For a reliable classifier, a few hundred images per group is a good starting point; for detection, the number of labelled boxes matters most. Not enough data? Start from a ready-made set in the Dataset Library.
No. VisionDock learns from data you have already labelled. For detection and localization you upload an annotated ZIP in YOLO, COCO or VOC format, which virtually every labelling tool exports. Classification only needs a folder per group, and multi-label or regression only needs a simple CSV.
It depends on dataset size and model. You see an estimated duration before a job starts, and progress is visible while training runs in the cloud.
After deployment you get a REST endpoint and an API key, and you call it with a file upload or base64 JSON. For detection you can pass confidence threshold, IoU, TTA, ensembling and SAHI parameters per request. For offline or edge scenarios, download the edge bundle and run it locally.
Your uploaded data and trained models belong only to your account and project; they are never added to the libraries or shared with other users. Access is controlled through verified accounts and API keys, which you can rotate or revoke at any time.
In short
Discovery, data checks, managed training and a live prediction API. No ML team, no servers to manage.