Multi-Task Learning for Dense Prediction Tasks: A Survey.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Deep learning significantly improved dense prediction tasks, typically by training separate models for each task.
• Multi-task learning (MTL) offers a more efficient approach by jointly learning multiple tasks using a shared representation.

Purpose of the Study:

• To provide a comprehensive overview of state-of-the-art deep learning methods for MTL in computer vision, focusing on dense prediction.
• To analyze MTL from both network architecture and optimization perspectives.
• To experimentally evaluate the effectiveness of various MTL strategies.

Main Methods:

• Surveying and categorizing recent MTL network architectures, discussing their strengths and weaknesses.
• Examining and comparing different optimization techniques for joint multi-task training.
• Conducting extensive experimental evaluations on dense prediction benchmarks.

Main Results:

• MTL approaches demonstrate potential for improved performance, reduced computation, and memory footprint compared to isolated task learning.
• Architectural choices and optimization strategies significantly impact MTL effectiveness.
• Empirical results validate the benefits of MTL across various dense prediction tasks.

Conclusions:

• Jointly learning dense prediction tasks via MTL offers significant advantages over training individual models.
• The choice of network architecture and optimization method is crucial for successful MTL implementation.
• This survey provides valuable insights for researchers and practitioners in applying MTL to computer vision problems.