2019/11/25 by Sergey Zakharov, Zakharov, Sergey, Wadim Kehl +5 · 4 citations
Engineering · Computer Science · #3D Shape Modeling and Analysis #Human Pose and Action Recognition #Advanced Neural Network Applications
paper · pdf · doi:10.48550/arxiv.1911.11288
We present an automatic annotation pipeline to recover 9D cuboids and 3D\nshapes from pre-trained off-the-shelf 2D detectors and sparse LIDAR data. Our\nautolabeling method solves an ill-posed inverse problem by considering learned\nshape priors and optimizing geometric and physical parameters. To address this\nchallenging problem, we apply a novel differentiable shape renderer to signed\ndistance fields (SDF), leveraged together with normalized object coordinate\nspaces (NOCS). Initially trained on synthetic data to predict shape and\ncoordinates, our method uses these predictions for projective and geometric\nalignment over real samples. Moreover, we also propose a curriculum learning\nstrategy, iteratively retraining on samples of increasing difficulty in\nsubsequent self-improving annotation rounds. Our experiments on the KITTI3D\ndataset show that we can recover a substantial amount of accurate cuboids, and\nthat these autolabels can be used to train 3D vehicle detectors with\nstate-of-the-art results.\n