Compare detection outputs against ground truth.
Load one labeled reference and two to four result files. Arena scores each run, identifies exact error frames, and exports the comparison without uploading the data.
Arena scores result files. It does not execute model weights.
Load detections
Accepted detection format
Use integer frame values and matching coordinate units in every file. JSON accepts a detection array, COCO annotations, a detections array, or frames with nested detections. CSV accepts x1, y1, x2, y2 or x, y, width, height. Confidence defaults to 1. Labels default to object.
Set the match rule
A prediction matches one ground-truth box when overlap meets this threshold.
Predictions below this score are excluded before matching.
Predictions are ordered by confidence, then matched once within the same frame and class to the available ground-truth box with the highest IoU. Scores aggregate every loaded frame.