We could further improve the performance with the data augmentation of NYUv2 defined in: [1] PAD-Net: Multi-Tasks Guided Prediction-and-D

Credit to Jialong Wu from https://github.com/lorenmt/mtan/issues/34.

SegNet MTAN

Dynamic Weight Average from `End-to-End Multi-Task Learning with Attention`. Modification of: https://github.com/lorenmt/mtan/blob/master/im2im_pr

TOWARDS IMPARTIAL MULTI-TASK LEARNING: https://openreview.net/pdf?id=IMPnRXEWpvr

Linear scalarization baseline L = sum_j w_j * l_j where l_j is the loss for task j and w_h

Based on the official implementation of: Multi-Task Learning as Multi-Objective Optimization Ozan Sener, Vladlen Koltun Neural Information Pro

Modification of: https://github.com/WeiChengTseng/Pytorch-PCGrad/blob/master/pcgrad.py @misc{Pytorch-PCGrad, author = {Wei-Cheng Tseng},

Random loss weighting: https://arxiv.org/pdf/2111.10603.pdf

Single task learning

Linear scalarization baseline L = sum_j w_j * l_j where l_j is the loss for task j and w_h

Implementation of `Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics` Source: https://github.com/yaringal/mul