Multi-Task-Learning-PyTorch

Why we included this project

If you work on dense prediction tasks like semantic segmentation, depth estimation, or surface normals, this repo gives you a working set of multi-task learning models to compare. It bundles several architectures and training strategies that share representations across tasks, tested on NYUD and PASCAL, and includes MTI-Net from ECCV 2020, which models cross-task interactions at multiple scales. The config-driven setup makes it easy to swap backbones or datasets once you adjust the path files. Because it accompanies a published survey on multi-task learning for dense prediction, it's a good starting point for reproducing paper results or building your own baseline. Just be ready for some manual setup and a specific PyTorch version; this is research code, not a plug-and-play library.