fiftyone-model-evaluation

To evaluate model predictions against ground truth in FiftyOne, you can use the Model Evaluation Panel for interactive assessment or the Python SDK for programmatic evaluation. Both methods support detections, classifications, segmentations, and regressions.

Available evaluation protocols include COCO, Open Images, and custom methods. These protocols allow you to compute metrics like mAP, precision, recall, and per-class metrics for object detection model predictions.

Yes, you can use the FiftyOne Python SDK to programmatically compute mAP, precision, recall, and per-class metrics. You need to specify prediction and ground-truth fields, an evaluation key, the method, and IoU threshold.

The best way to compare two object detection models is by computing mAP, precision, recall, and per-class metrics using unique evaluation keys for each model. You can then inspect failures via evaluation patches to identify differences.

Yes, you need correctly named prediction and ground-truth fields in your dataset to run model evaluation. You must also provide unique evaluation keys to manage and store the results of each evaluation run separately.

The FiftyOne evaluation plugin is required when you need specific evaluation protocols or operators not included in the core package. It extends functionality for interactive and programmatic model evaluation tasks.