[Paper Review] Multi-Task Learning for Dense Prediction Tasks: A Survey

Simon Vandenhende, Stamatios Georgoulis+4 more|Lirias (KU Leuven)|Apr 28, 2020

Domain Adaptation and Few-Shot Learning70 references53 citations

TL;DR

A comprehensive survey of deep multi-task learning for pixel-level vision tasks, detailing architectures, optimization methods, and empirical evaluations.

ABSTRACT

With the advent of deep learning, many dense prediction tasks, i.e. tasks that produce pixel-level predictions, have seen significant performance improvements. The typical approach is to learn these tasks in isolation, that is, a separate neural network is trained for each individual task. Yet, recent multi-task learning (MTL) techniques have shown promising results w.r.t. performance, computations and/or memory footprint, by jointly tackling multiple tasks through a learned shared representation. In this survey, we provide a well-rounded view on state-of-the-art deep learning approaches for MTL in computer vision, explicitly emphasizing on dense prediction tasks. Our contributions concern the following. First, we consider MTL from a network architecture point-of-view. We include an extensive overview and discuss the advantages/disadvantages of recent popular MTL models. Second, we examine various optimization methods to tackle the joint learning of multiple tasks. We summarize the qualitative elements of these works and explore their commonalities and differences. Finally, we provide an extensive experimental evaluation across a variety of dense prediction benchmarks to examine the pros and cons of the different methods, including both architectural and optimization based strategies.

Motivation & Objective

• Provide a unified view of state-of-the-art MTL techniques for dense prediction tasks in computer vision.
• Classify architectures by where task interactions occur (encoder-focused vs decoder-focused) and summarize their pros/cons.
• Review optimization strategies for balancing multiple tasks during training.
• Present extensive experimental comparisons across benchmarks to evaluate architectural and optimization approaches.
• Share publicly available code to facilitate adoption of reviewed techniques.

Proposed method

• Taxonomy of MTL architectures based on interaction location (encoder-focused vs decoder-focused).
• Survey of encoder-focused methods (e.g., cross-stitch, NDDR-CNN, MTAN, branched MTL).
• Survey of decoder-focused methods (e.g., PAD-Net, PAP-Net, MTI-Net, JTRL, MTI-Net).
• Discussion of optimization strategies for task balancing (e.g., fixed uncertainty, GradNorm, DWA, multi-objective optimization).
• Comprehensive experimental evaluation across diverse dense-prediction benchmarks to compare methods.
• Provision of public code to ease adoption (GitHub link).

Experimental results

Research questions

• RQ1 What architectural choices for multi-task learning best leverage shared representations for dense prediction tasks?
• RQ2 How do encoder-focused and decoder-focused MTL architectures compare in performance and efficiency?
• RQ3 Which optimization strategies most effectively balance multiple tasks during training?
• RQ4 How do different MTL approaches perform across diverse dense-prediction benchmarks in a fair apples-to-apples setting?

Key findings

• Encoder-focused and decoder-focused architectures offer complementary strengths, with decoder-focused approaches often providing richer cross-task interactions via multi-scale distillation.
• Soft parameter sharing methods (e.g., cross-stitch, NDDR-CNN, MTAN) address feature sharing but face scalability challenges as task count grows.
• Decoder-focused models that perform multi-modal or multi-scale distillation (e.g., PAD-Net, PAP-Net, MTI-Net) capture richer task relationships than encoder-only sharing alone.
• Branched MTL networks automate task grouping and sharing to balance performance and computation under budgets, showing end-to-end optimization capabilities.
• Optimization strategies for task balancing (e.g., fixed uncertainty, GradNorm, DWA, multi-objective optimization) are crucial to avoid dominance by any single task and improve joint performance.
• Extensive experiments across benchmarks provide apples-to-apples insights into the relative strengths and trade-offs of architectural and optimization choices.