Putting Current State of the art Object Detectors to the Test: Towards Industry Applicable Leather Surface Defect Detection

Automated leather defect classification has gained a lot of attention in recent years with the advancement of automation in the leather industry. In recent years a plethora of new techniques have been presented which aim to improve the automated defect classification pipeline. However, not much attention is given to the important task of automated defect detection along with categorization. The aim of this work is to evaluate the current state object detectors and their systematically adapted variants for the task of leather defect detection. A major goal of this study is to provide recommendations for model identification/selection and design of new detection methods for the task of leather defect detection. Important findings of this study are that multistage detectors better generalize to difficult defects of varying characteristics. Also, shallower backbones were found to be more accurate and are well-suited to problems where a relatively limited amount of data is available for training.