🚀 Quick-start Guide

New to Ocular Foundry’s model training capabilities? Follow this five-minute path to train and evaluate your first custom model. Once you are comfortable, circle back to the advanced pages for deep dives.

1. Requirements

ItemWhy you need it
At least one VersionTraining uses the frames and annotations in a specific version.
Member role permissionsYou must have member rights to launch training jobs.

2. Open the Training tab

  1. Navigate to the Versions tab in a project.
  2. Select the version you wish to use to train the model
  3. Click on the Train in the right-hand corner of that version’s details page. You will see the following popup to configure and start a training run:

3. Adjust Configuration Settings

Choose the model type, model variant, number of epochs, and image size.

SettingDefaultRationale
Model VariantmBest blend of speed and accuracy.
Epochs10Enough to verify the workflow.
Image Size640Detects small objects yet fits in 8 GB VRAM.

YOLOv11 supports detection, segmentation, pose, classification, and oriented bounding boxes (OBB). Ocular currently supports detection and segmentation models for object detection or segmentation project types, respectively; other types are on the roadmap.

4. Click Create Model

You will be redirected to the “Models” tab, where the run appears instantly in Model Training History with a processing 🔵 status. When the training run has completed, it will have a competed 🟢 status.

5. Observe Performance Metrics

Performance Metrics and loss curves will be available on the model details page after the training run has completed 🟢 . You will see the precision, recall, and mAP metrics over epochs, the confusion matrix, and previews of validation set data, and more.

That’s it — you now own a custom detector fine-tuned on your data.

Next up: