Object Detection in Indian Food Platters using Transfer Learning with YOLOv4

Object detection is a well-known problem in computer vision. Despite this, its usage and pervasiveness in the traditional Indian food dishes has been limited. Particularly, recognizing Indian food dishes present in a single photo is challenging due to three reasons: 1. Lack of annotated Indian food datasets 2. Non-distinct boundaries between the dishes 3. High intra-class variation. We solve these issues by providing a comprehensively labelled Indian food dataset- IndianFood10, which contains 10 food classes that appear frequently in a staple Indian meal and using transfer learning with YOLOv4 object detector model. Our model is able to achieve an overall mAP score of 91.8% and f1-score of 0.90 for our 10 class dataset. We also provide an extension of our 10 class dataset- IndianFood20, which contains 10 more traditional Indian food classes.