Diagnose Label Errors for 3D Object Classification

Abstract

3D object classification is widely used in many real-life scenarios and has recently become a popular research area. Meanwhile, 3D datasets tend to be larger and more complex, increasing the difficulty of labeling. Since label errors in test data affect the accuracy of model evaluation and analysis, this study proposes a multi-view post-training label refinery protocol to clean the label errors in test data. Experimental results on ModelNet40 show that 4.58% of test samples are selected as label-error candidates and 29.2% of them are relabeled. On this basis, this study analyzes the weakness of ModelNet40 and provides suggestions for future dataset construction.