ESeNet-D : Efficient Semantic Segmentation for RGB-Depth Food Images

Abstract

Segmenting food images is crucial for health-related tasks such as estimating dietary nutrients and offering personalized diet recommendations. However, developing segmentation model demands abundant annotated images and substantial computing resources. Therefore, there is a need for high-performing yet efficient models in terms of size, training, and inference times. Existing research predominantly focuses on Asian or Western cuisines and RGB-D segmentation for food images remains underexplored due to limited datasets. In this paper, we propose an efficient RGB-D segmentation model that surpasses reference models in both RGB and RGB-D domains while utilizing fewer parameters and floating-point operations. Our model is evaluated on the CamerFood15 dataset, an updated African food image dataset, achieving an mIoU of 84.58% and an Overall Pixel Accuracy of 95.61%.