Occlusion resilient similar-colored separable food item instance segmentation

Abstract

The task of recognizing non-Western and non- Chinese food items as well as accurately segmenting food item instances is a seldom researched and challenging task in the field of Computer Vision. Food items such as Sri Lankan short eats snacks have high inter-class visual similarity, mainly in terms of color and the fact that food images are highly prone to occlusion or item overlap where a portion of an object is hidden from sight. Existing databases are few and synthetic and current systems do not handle food item occlusion. In this paper a novel Sri Lankan short eats food item instance segmentation and amodal completion approach is introduced as well as two novel datasets for Sri Lankan short eats instance segmentation and amodal instance segmentation. The proposed method shows model performance improvements up to 88.4% mAP in Instance Segmentation and up to 90% mIoU in Amodal Completion, as well as the advantage of real-time inference in less than 1.7 seconds per frame.