Detecting Furniture in Images Based on Neural Network Models

Sammanfattning: Object detection is a challenging task that locates objects within an image or video and thenallows us to count and then track those objects. It is applied in a variety of ways, such as crowd counting, self-driving vehicles, video surveillance, face identification, and anomaly detection, and is utilized in a wide range of industries, including retail, sport, healthcare, marketing, interior design, agriculture, construction, and recreation. Despite the tremendous advancements made in this area and the impressive capabilities of computer vision, object detection is still a challenging process that faces challenges its application, such as the variety of objects and algorithm processing speed. Finding and classifying individual objects within an image used to be a very challenging task in the past, however, deep learning has made it possible for algorithms to easily and rapidly recognize the content of images, opening up new ways for the study and analysis of visual data in a variety of fields.The recent top-performing deep learning model families on this task include R-CNN, YOLO, and SSD. This master thesis implements and examines two state-of-the-art neural network models from these families, i.e., Faster R-CNN and YOLOv3, and compares their performance on the whole dataset and every single class. The results of this study show that these two models have similar average precision on the whole dataset, but Faster R-CNN works better in detecting bed and couch from images while YOLOv3 is better at recognizing tables. It is also found that YOLOv3 has a shorter training time and produces results faster because of its simpler architecture. This study ends with a discussion on the limitations of models, such as insufficient training epochs and inappropriate data augmentation.