What Is Portal?
Portal is Datature's open-source tool for inspecting model predictions on images and videos. After training a model, you need to see what it actually does on real data: where it draws boxes, how confident the predictions are, and where it fails. Portal provides that visual inspection layer without writing any code.
What This Tutorial Covers
Datature walks through using Portal to inspect computer vision model outputs:
- Loading a trained model and a set of test images into Portal
- Running inference and viewing predictions overlaid on each image
- Adjusting confidence thresholds to filter weak predictions
- Comparing predictions across different model versions
- Identifying failure patterns: missed detections, false positives, and confused classes
The tutorial takes about five minutes and applies to object detection, segmentation, and classification models.
Why Visual Inspection Matters
Metrics like mAP and precision give you numbers, but they do not show you why the model fails. A 90% mAP model might consistently miss small objects in dark corners, or it might hallucinate detections on reflective surfaces. Portal lets you spot these patterns by scrolling through predictions one image at a time. This kind of inspection often reveals annotation errors in the training data too.
For more on Portal, read Inspect Model Inferences on Images and Videos with Portal. To interpret the training metrics themselves, see How to Interpret Training Graphs.
