Ultra: Unsupervised Cross-Task Optimization for Reliable Restoration Segmentation Collaboration under Adverse Weather

TL;DR

Ultra framework achieves 73.0% mIoU on ACDC using CTDN and CMIL for unsupervised cross-task optimization under adverse weather.

cs.CV 🔴 Advanced 2026-08-17 31 views
Shiqin Wang Zhiqian Li Haoyuan Du Junming Chen Jiayuan Li Tianrun Xu Haoyang Chen
unsupervised learning semantic segmentation image restoration cross-task optimization causal inference

Key Findings

Methodology

Ultra integrates CTDN and CMIL. CTDN generates optimization directions via bidirectional task adaptation and selects optimal directions using Nash bargaining. CMIL uses causal interventions to suppress error propagation.

Key Results

  • Achieved 73.0% mIoU on ACDC with HRDA backbone, surpassing the best baseline by 0.4%.
  • Reached 59.3% mIoU on nighttime driving in Cityscapes→Dark Zurich, outperforming state-of-the-art.
  • Exceeded existing UDA-ASS restoration methods in unsupervised image restoration tasks.

Significance

Ultra addresses cross-task collaboration challenges under adverse weather, offering a reliable optimization framework that enhances segmentation and restoration performance while extending to tasks like object detection.

Technical Contribution

Ultra introduces CTDN for direction selection via Nash bargaining and CMIL for causal inference, establishing a novel paradigm for cross-task collaboration and reducing error propagation.

Novelty

Ultra is the first to apply causal interventions in unsupervised cross-task optimization, addressing uncertainty and hallucination issues in prior methods.

Limitations

  • Performance may degrade under extreme weather conditions like heavy fog.
  • High computational complexity increases training time.

Future Work

Future work could explore Ultra's extension to multi-task scenarios (e.g., detection, tracking) and improve computational efficiency for real-time applications.

AI Executive Summary

Semantic segmentation and image restoration under adverse weather are critical for applications like autonomous driving. Existing methods assume mutual guidance between tasks, but this often leads to error propagation in unsupervised settings.

The Ultra framework introduces CTDN and CMIL to enable reliable cross-task collaboration. CTDN generates optimization directions through bidirectional task adaptation and selects optimal ones using Nash bargaining. CMIL evaluates the reliability of cross-task information transfer using causal interventions, suppressing error propagation. Experiments demonstrate state-of-the-art performance on ACDC and Dark Zurich datasets.

Beyond segmentation and restoration, Ultra shows potential for tasks like object detection. However, its performance under extreme weather and computational efficiency require further optimization. Future work could explore its application to multi-task scenarios and real-time systems.

Deep Analysis

Background

Semantic segmentation and image restoration are essential for autonomous driving, but annotating adverse weather data is costly. Unsupervised domain adaptation (UDA) methods transfer labeled clear-weather data to adverse-weather scenarios, reducing annotation needs.

Core Problem

Existing UDA methods face challenges in cross-task collaboration, including ambiguous optimization directions and error propagation, leading to degraded performance. Reliable cross-task optimization remains a key challenge.

Innovation

Ultra introduces CTDN for direction selection and CMIL for causal interventions, addressing uncertainty and error propagation. This is the first framework to apply causal inference to cross-task collaboration.

Methodology

  • �� CTDN: Generates candidate optimization directions via bidirectional task adaptation and selects optimal ones using Nash bargaining.
  • �� CMIL: Evaluates cross-task information transfer reliability using causal interventions to suppress error propagation.
  • �� Training: Conducted on Cityscapes and ACDC datasets with mixed-domain optimization for restoration and segmentation.

Experiments

Experiments were conducted on Cityscapes→ACDC and Cityscapes→Dark Zurich tasks using DeepLabV2, DAFormer, and HRDA backbones. Metrics include mIoU, with additional ablation studies and restoration performance comparisons.

Results

Ultra achieved 73.0% mIoU on ACDC, surpassing prior methods. On Dark Zurich, it reached 59.3% mIoU for nighttime driving. It also outperformed existing methods in unsupervised image restoration.

Applications

Ultra is applicable to autonomous driving and surveillance systems requiring robust segmentation and restoration under adverse weather. Its extensibility supports tasks like object detection.

Limitations & Outlook

Performance under extreme weather conditions needs improvement. High computational complexity may limit real-time applications.

Plain Language Accessible to non-experts

Imagine you're in a foggy room trying to find the exit. Restoration is like cleaning the windows to see better, while segmentation is like a map showing the exit. Ultra ensures these two processes work together, avoiding mistakes caused by misleading visuals.

ELI14 Explained like you're 14

Think of playing a game in a foggy world. Restoration clears the fog so you can see, and segmentation marks important spots on your map. Ultra is like a smart helper making sure these tools work together without messing things up!

Glossary

Unsupervised Domain Adaptation (UDA)

A transfer learning method that adapts knowledge from labeled source domains to unlabeled target domains.

Used to transfer clear-weather annotations to adverse-weather scenarios.

CTDN

Cross-Task Direction Negotiation module for generating and selecting optimization directions.

Used for direction selection between restoration and segmentation tasks.

CMIL

Causal Mutual Intervention Learning module for evaluating and suppressing error propagation.

Used to assess cross-task information transfer reliability.

Nash Bargaining

A game-theoretic method for finding optimal balance between multiple objectives.

Used in CTDN for optimization direction selection.

mIoU

Mean Intersection-over-Union, a standard metric for semantic segmentation accuracy.

Used to evaluate segmentation performance.

Open Questions Unanswered questions from this research

  • 1 How to further improve performance under extreme weather conditions?
  • 2 How to reduce computational complexity for real-time applications?

Applications

Immediate Applications

Autonomous Driving

Enhances environmental perception under adverse weather, improving driving safety.

Surveillance Systems

Improves object detection in low-light or adverse weather conditions.

Long-term Vision

Multi-task Collaboration

Extends to tasks like detection and tracking, enabling comprehensive visual perception systems.

Abstract

Unsupervised Domain Adaptation for Adverse Weather Semantic Segmentation (UDA-ASS) aims to transfer semantic knowledge from labeled normal-weather images to unlabeled adverse environments. Existing approaches implicitly assume that restoration and segmentation provide mutually beneficial guidance. However, under severe degradation and without target-domain supervision, the validity of cross-task optimization directions becomes fundamentally unidentifiable, leading to hallucination-driven error propagation. In this work, we propose a novel Unsupervised Restoration-Segmentation Collaborative Learning Framework (Ultra), which reframes cross-task interaction as direction selection under uncertainty and causal effect estimation, enabling reliable collaboration through candidate direction generation and intervention-based filtering. In detail, we propose CTDN and CMIL. The former exploits complementary visual structures and semantic information to generate candidate optimization directions and performs cooperative direction selection between restoration and segmentation. The latter reformulates cross-task information transfer from correlation-based propagation into causal effect assessment, suppressing hallucination propagation. Extensive experiments on three widely used UDA-ASS benchmarks demonstrate state-of-the-art segmentation performance. Beyond segmentation, our framework achieves better unsupervised restoration results than existing UDA-ASS restoration methods and generalizes to unsupervised restoration and object detection collaboration tasks. Code and models will be available at https://github.com/Wang-Shiqin/Ultra.

cs.CV