Demystifying IoU score prediction in 3D LIDAR-based object detectors

In this paper, we aim to identify how various architectural choices in two-stage one-to-many unimodal LIDAR 3D detectors influence the calibration of their IoU score (S<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">IoU</inf>) prediction branch. We select a relevant family of freely available networks, and extensively test their output scores w.r.t. S<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">IoU</inf> regression, and score-based ordering of predicted bounding box (BBox). We observe that the interpretation of predicted localization score (Ŝ<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">IoU</inf>) as a confidence score is poorly justified, that geometric features seem to be of importance when it comes to accurately predicting S<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">IoU</inf>, and that accurate ranking over the whole score range seems to be favored by the use of a more balanced, in terms of target metric values, distribution of BBoxes in training the score.