Research map / Label-efficient and robust learning / Recognition and long-tail

Learning Real-World Image De-weathering with Imperfect Supervision

AAAI 2024 · Label-efficient and robust learning › Recognition and long-tail · approach: Group-robust classification · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a Consistent Label Constructor (CLC) that uses multi-frame inputs to produce pseudo-labels consistent with degraded inputs, and an Information Allocation Strategy (IAS) that jointly supervises de-weathering with pseudo-labels (L1) and original labels (rain-robust contrastive plus sliced Wasserstein losses) to handle imperfect real-world supervision.

Problem

Real-world de-weathering datasets captured sequentially contain illumination, position, and texture inconsistencies between degraded inputs and ground-truth labels, which degrade supervised training of de-weathering networks.

Approach

Use information bottleneck theory to build a pseudo-label from a multi-frame label constructor that stays consistent with the input while removing most degradation, then combine it with the original label via a loss allocation that uses L1 for the pseudo-label and feature/distribution-level losses (robust to misalignment and color shifts) for the original label.

Key results

• Ours-Restormer improves overall PSNR from 21.51 to 22.17 dB on GT-Rain-Snow, and Ours-RainRobust improves 22.01 to 22.40 dB.
• On WeatherStream, Ours-Restormer and Ours-RainRobust gain 0.21 dB and 0.24 dB PSNR respectively over their baselines.
• Ablation: using only the pseudo-label raises RainRobust overall PSNR from 22.01 to 22.18 dB; combining with the original label reaches 22.40 dB.

Measured effects reported in the paper

Surprising findings

• Pseudo-labels alone as supervision outperform the original imperfect labels (22.18 vs 22.01 dB PSNR), though combining both is best.

Limitations

• Gains are modest in absolute PSNR (about 0.2 to 0.7 dB).
• CLC requires multiple adjacent frames, which limits use to video-derived datasets.
• Pseudo-label still trades off degradation removal against consistency.

Experimental setup

Datasets: GT-Rain-Snow WeatherStream Internet-Data

Benchmarks: GT-Rain-Snow test split (45 scenes) Out-of-distribution rain/snow/fog images

Metrics: PSNR SSIM PI CLIP-IQA

Baselines: Restormer RainRobust Guided Filter Optical Flow Misalign-Tolerate l1

Builds on: Restormer RainRobust information bottleneck theory InfoNCE

Closest papers by meaning

• Continuous Adverse Weather Removal via Degradation-Aware Distillation (CVPR 2025 · similarity 0.85)
• Real-World Adverse Weather Image Restoration via Dual-Level Reinforcement Learning with High-Quality Cold Start (NeurIPS 2025 · similarity 0.85)
• SnowMaster: Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization (CVPR 2025 · similarity 0.85)
• CoA: Towards Real Image Dehazing via Compression-and-Adaptation (CVPR 2025 · similarity 0.84)
• Semantic Segmentation in Multiple Adverse Weather Conditions with Domain Knowledge Retention (AAAI 2024 · similarity 0.84)
• AWRaCLe: All-Weather Image Restoration Using Visual In-Context Learning (AAAI 2025 · similarity 0.84)