Comparison between 2D Classification and 3D Classification based on Deep Learning

• Published in: 2024 3rd International Conference on Embedded Systems and Artificial Intelligence (ESAI) pp. 1 - 6
• Main Authors: Khalfi, Zouhir El, Hazzat, Soulaiman El, Yakhlef, Majid Ben
• Format: Conference Proceeding
• Language: English
• Published: IEEE 19.12.2024
• Subjects: 2D images data and Convolution Neuronal networks; 3D points cloud data; Accuracy; Classification; Classification algorithms; Convolution; Data models; Deep learning; Natural language processing; Point cloud compression; Solid modeling; Three-dimensional displays; Timing
• DOI: 10.1109/ESAI62891.2024.10913611

The rise in processing power of computers and the availability of a considerable amount of data have allowed the overlap of several machine learning approaches. This has led to breakthroughs on many deep learning tasks as speech recognition, natural language processing (NLP), object detection and image analysis. Thus, many fields as autonomous driving, industrial quality processes and mobile robotic have released multiples researches in deep learning algorithmic topics. In this paper, we have made a comparison, in terms of accuracy and process timing, between classifications based on 3D points cloud and 2D images for the same input dataset in raw format, and generated basic outlines of both deep learning methods using algorithms based on convolution networks. We discovered that 3D points cloud based classification model was slow on training stage than 2D images model timing due to the 3D complexity data representation. Besides that, classification neural networks based on 3D points cloud data performed better on accuracy than the networks based on 2D images datasets. It is to highlight that 3D based systems are so critical in terms of data processing compared to 2D images due to: Data format as points cloud, meshes and volumes, the hardware resources requirement which is higher due to the third added dimension. Finally, we hypothesize that systems which would combine both of these data geometry and networks architectures would return accuracy better and low processing time wise.