Reasoning and Learning a Perceptual Metric for Self-Training of Reflective Objects in Bin-Picking With a Low-Cost Camera

Bin-picking of metal objects based on low-cost RGBD cameras may suffer errors due to sparse depth information and reflective part texture, leading to a need for manual labeling. To reduce the need for human intervention, we propose a framework consisting of a metric learning stage and a self-training stage. Specifically, to automate the handling of the data captured by a low-cost camera (LC), a multi-object pose reasoning (MoPR) algorithm is proposed which optimizes the data with depth, collision and boundary constraints. To improve the accuracy of the retrieved pseudo labels, we generalize the traditional Bayesian Gaussian Mixture Model (BGMM) to the configuration manifolds and propose Symmetry-aware Lie-group based BGMM (SaL-BGMM). Moreover, to solve the ambiguity issue of perceptual metric learning for multiple objects, we propose a weighted ranking information noise contrastive estimation (WR-InfoNCE) loss. Experiments show that our algorithm outperforms related popular works both on ROBI and our proposed Self-ROBI dataset.