The conversion of 3D shapes from 2D observations is a significant challenge within the computer vision field with wide applications in robotics, medical imaging, and industrial design. Traditional methods are prone to occlusion, noise, and ambiguity in mapping, thereby rendering them ineffective in real-world applications. This research work introduces an AI-based paradigm employing a Convolutional Neural Network (CNN), infused with a new metaheuristic optimization technique, the advanced Hippopotamus Optimizer (AHO). The encoder-decoder CNN architecture undergoes training, making it capable of understanding the translation from a 2D image to an accurate 3D structure, and the AHO enhances optimization of parameters through balancing exploration and exploitation towards fast convergence and greater accuracy production. Testing on the ShapeNet dataset confirms that this framework outperforms conventional CNNs trained with gradient descent and alternative metaheuristic optimizers under single-view, multi-view, and noisy scenarios. Noteworthy improvement in reconstruction accuracy and robustness, proving AHO’s potential as a very capable optimization tool for deep learning in challenging vision applications.
Li et al. (Tue,) studied this question.