Accurate 3D reconstruction is essential for robotic applications that require interaction with complex structures. Traditional methods relying on sensor-based point cloud data often struggle with noise and occlusion, leading to incomplete or imprecise reconstructions. This thesis explores and compares two approaches to 3D reconstruction: Image to 3D reconstruction, which predicts a full 3D model from a single image, and 3D to 3D reconstruction, which reconstructs a complete object from a partial 3D input.

Image to 3D reconstruction is inherently challenging due to the difficulty of inferring missing depth information from a single viewpoint. Our experiments demonstrate that an encoder-decoder architecture trained on 3D skeleton data can reconstruct meaningful structures, but the accuracy is highly sensitive to thresholding. At lower confidence thresholds, the model captures more details but introduces false positives, while higher thresholds suppress errors at the cost of missing finer structures.

3D to 3D reconstruction, by contrast, benefits from richer spatial information provided by an incomplete 3D input. The results show that this method achieves consistently higher F1-scores and IoU across all thresholds, making it significantly more robust to occlusions. Architectural choices also play a crucial role; models with skip connections outperform deeper or more complex variations, highlighting the importance of feature retention in volumetric reconstruction. Additionally, our study of loss functions confirms that Weighted Binary Cross-Entropy (WBCE) provides the best reconstruction accuracy at lower thresholds, whereas F1 Loss offers more stable performance across varying levels of occlusion.

The findings of this thesis suggest that 3D to 3D reconstruction is the preferred method for scenarios with significant occlusions, while Image to 3D reconstruction remains viable in settings where only 2D inputs are available. Future work should explore hybrid models that combine 2D and 3D data, develop adaptive thresholding techniques to improve reconstruction confidence, and conduct real-world validation to bridge the gap between experimental and practical applications.