Disentangling Monocular 3D Object Detection

In this paper we propose an approach for monocular 3D object detection from asingle RGB image, which leverages a novel disentangling transformation for 2Dand 3D detection losses and a novel, self-supervised confidence score for 3Dbounding boxes. Our proposed loss disentanglement has the twofold advantage ofsimplifying the training dynamics in the presence of losses with complexinteractions of parameters, and sidestepping the issue of balancing independentregression terms. Our solution overcomes these issues by isolating thecontribution made by groups of parameters to a given loss, without changing itsnature. We further apply loss disentanglement to another novel, signedIntersection-over-Union criterion-driven loss for improving 2D detectionresults. Besides our methodological innovations, we critically review the APmetric used in KITTI3D, which emerged as the most important dataset forcomparing 3D detection results. We identify and resolve a flaw in the 11-pointinterpolated AP metric, affecting all previously published detection resultsand particularly biases the results of monocular 3D detection. We provideextensive experimental evaluations and ablation studies on the KITTI3D andnuScenes datasets, setting new state-of-the-art results on object category carby large margins.