An efficient technique for real-time transformation of 2D to 3D images with GPU using CUDA programming

Abstract

This thesis presents a complete 2D to 3D reconstruction system designed to run reliably on a low computational powered PC, where GPU memory, host memory, and disk bandwidth impose strict constraints. The pipeline begins with large-scale synthetic data generation from ShapeNet models, producing aligned RGB and depth observations for supervised learning. A ResUNet18 based monocular depth network is trained in LibTorch using a mask-aware objective to promote numerical stability and reduce invalid-depth regions in the predicted maps. To ensure continuous training without data starvation under limited resources, the system is implemented a producer consumer scheduling system design: a producer renders and stages batches to fast local storage, consumers stream and pre-process shards into the training loop, and a destroyer reclaims storage deterministically once a batch is fully consumed. This design bounds disk usage, prevents host RAM accumulation, and decouples rendering from training so the GPU remains saturated even when CPU-side work fluctuates. After inference, predicted camera-centric depth is lifted into explicit 3D geometry using CUDA-accelerated reconstruction, enabling dense point cloud and grid-mesh generation at image resolution with minimal overhead. The system is evaluated using runtime traces (GPU utilization, GPU/host memory, and CPU load) alongside standard depth estimation metrics aggregated across training batches, demonstrating sustained execution, stable memory behavior, and reconstruction-ready depth quality on resource-constrained hardware.