Multitask Learning : A Unified Multi-task Learning for Semantic Segmentation andObject Detection in 3D Point Cloud

3D scene understanding plays an important role in applications such as robotics and autonomous driving. Two key tasks in this area are semantic segmentation and object detection in 3D point clouds. These tasks are usually solved by separate networks, which leads to repeated feature extraction, higher computational cost, and longer inference time. This makes it difficult to use such systems in real-time environments. In this thesis, we present a unified multi-task learning (MTL) framework that integrates two established models, SparseConvNet U-Net for semantic segmentation and VoteNet for object detection. The framework allows for knowledge exchange across the tasks, with a segmentation-guided Farthest Point Sampling (FPS) point selection mechanism to focus detection on object-relevant regions. This approach produces improved predictions while also reducing redundant computations. Evaluated on the ScanNet Dataset, our framework achieved 60.0%mIoU for semantic segmentation and 37.8% mAP for object detection, outperforming its own baselines (55.7% mIoU and 33.4% mAP, respectively). Furthermore, it resulted in efficiency and computational complexity gains, including 2.9% fewer parameters, 2.7 ms faster inference, and 41% fewer MAC operations compared to combined single-task baselines. Qualitative results further confirm that combining these tasks yields a coherent scene understanding. This work demonstrates that MTL not only improves task accuracy but also offers an efficient and computationally cheaper solution for real-time 3D vision tasks, making it suitable for resource-constrained environments.