Deep Learning-based 3D Reconstructed Objects via Generative Models and Octree Structure

Recent advancements in intelligent systems and innovative technology have revolutionized data processing, analysis, and reconstruction. Machine learning and artificial intelligence breakthroughs now enable the restoration of missing or lost information from datasets. However, reconstructing data lost during processes like image acquisition or dimension reduction remains a significant challenge. This study introduces a method that systematically captures 2D object images and predicts voxel-based 3D representations to reconstruct the objects. Using a multi-layer encoder and decoder, the image's depth is estimated and combined with the original image to determine the object's 3D shape. EfficientNet and an octree-based technique generate voxels from the 3D shape representation. The proposed model is evaluated on three benchmark datasets, comparing against state-of-the-art methods. Performance is assessed using evaluation metrics like Chamfer distance (CD), earth mover's distance (EMD), and intersection over union (IoU). Results indicate mean Chamfer distances of 0.00387, 0.00317, and 0.00102 for ShapeNetCore, Pascal3D+, and Pix3D, respectively. Mean IoU values of 5.58, 6.13, and 6.72 are achieved on the mentioned benchmark datasets.