ConvMTL: Multi-task Learning via Self-supervised Learning for Simultaneous Dense Predictions

Perception systems in autonomous vehicles are required to perform multiple scene-understanding tasks under tight constraints of latency and power. Single-task neural networks can become unscalable when the number of tasks increases in the perception stack. Multi-task learning has been shown to improve parameter efficiency and enable models to learn more generalizable task representations compared to single-task neural networks. This work explores a novel convolutional multi-task neural network architecture that simultaneously performs two dense prediction tasks, semantic segmentation and depth estimation. A self-supervised ResNet-50 backbone is used as the basis of the proposed network, along with a multi-scale feature fusion module and a dense decoder. The model uses a simple weighted loss function with an informed search algorithm identifying the optimal parameters. The performance of the proposed model on the segmentation task is assessed using the mean Intersection of Union (mIoU) and pixel accuracy. In contrast, absolute and relative errors assess the depth estimation task. The obtained results for segmentation and depth estimation are mIoU of 73.81%, pixel accuracy of 93.52%, an absolute error of 0.130, and a relative error of 29.05. The model's performance is comparable to existing multitask algorithms on the Cityscapes dataset, using only 2975 training samples.