Most of the statistics are a function of the overlap requirement. When we have a potential overlap between a ground truth aand a detected box, there is the question of whether they are a correct pairing, or if the detected box is the same as the ground truth box. We have not done a deep correspondence analysis. Instead, we compute the ratio (area of the intersection of the boxes)/(area of the ground truth box). This is the overlap. A match is determined to occur if the the overlap is at least as big as the Overlap Requirement. The bigger the overlap requirement, the more the boxes are required to overlap. hence performance usually declines as the requirement reaches 100%.

• True Detections : the percent of ground truth boxes that have an overlapping data box
• Missed Detections : the percent of ground truth boxes that do not have an overlapping data box
• False Detections : the number of detected boxes that do not overlap a ground truth box, divided by the number of ground truth boxes
• True vs. False Detections : plot of true vs false detection as a function of Overlap Requirement
• Position Detection Errors : average pixels between center of ground truth and detected box, for true detections
• Size Detection Errors : average ratio of area differece to ground truth box area, for true detections
• Orientation Detection Errors : average degrees difference in axis orientations of ground truth and detected boxes, for true detections
• Time Lag Entry : average frames time difference between first ground truth minus detected frame, for true detections
• Time Lag Exit : average frames time difference between last ground truth minus detected frame, for true detections
• Dropped Frame Rate: average number of frames lost in the middle of a track, for true detections