# Heuristic Self-Paced Learning for Domain Adaptive Semantic Segmentation under Adverse Conditions

arXiv: https://arxiv.org/abs/2603.24322 · HTML: https://mlsanity.com/paper/2603.24322/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 21 of its 37 references that are in the map.
- First posted on arXiv 25 Mar 2026.

## Builds on (most cited first, 21 of 21)

- [DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs](https://mlsanity.com/paper/1606.00915/) (2016)
- [DAFormer: Improving Network Architectures and Training Strategies for Domain-Adaptive Semantic Segmentation](https://mlsanity.com/paper/2111.14887/) (2021)
- [HRDA: Context-Aware High-Resolution Domain-Adaptive Semantic Segmentation](https://mlsanity.com/paper/2204.13132/) (2022)
- [Refign: Align and Refine for Adaptation of Semantic Segmentation to Adverse Conditions](https://mlsanity.com/paper/2207.06825/) (2022)
- [Both Style and Fog Matter: Cumulative Domain Adaptation for Semantic Foggy Scene Understanding](https://mlsanity.com/paper/2112.00484/) (2021)
- [CSCL: Critical Semantic-Consistent Learning for Unsupervised Domain Adaptation](https://mlsanity.com/paper/2008.10464/) (2020)
- [Exploring Image Enhancement for Salient Object Detection in Low Light Images](https://mlsanity.com/paper/2007.16124/) (2020)
- [CoDA: Instructive Chain-of-Domain Adaptation with Severity-Aware Visual Prompt Tuning](https://mlsanity.com/paper/2403.17369/) (2024)
- [Hardness Sampling for Self-Training Based Transductive Zero-Shot Learning](https://mlsanity.com/paper/2106.00264/) (2021)
- [Contrastive Model Adaptation for Cross-Condition Robustness in Semantic Segmentation](https://mlsanity.com/paper/2303.05194/) (2023)
- [A Curriculum-style Self-training Approach for Source-Free Semantic Segmentation](https://mlsanity.com/paper/2106.11653/) (2021)
- [VBLC: Visibility Boosting and Logit-Constraint Learning for Domain Adaptive Semantic Segmentation under Adverse Conditions](https://mlsanity.com/paper/2211.12256/) (2022)
- [Condition-Invariant Semantic Segmentation](https://mlsanity.com/paper/2305.17349/) (2023)
- [Informative Data Mining for One-Shot Cross-Domain Semantic Segmentation](https://mlsanity.com/paper/2309.14241/) (2023)
- [Instance-Warp: Saliency Guided Image Warping for Unsupervised Domain Adaptation](https://mlsanity.com/paper/2403.12712/) (2024)
- [FREST: Feature RESToration for Semantic Segmentation under Multiple Adverse Conditions](https://mlsanity.com/paper/2407.13437/) (2024)
- [CoPT: Unsupervised Domain Adaptive Segmentation using Domain-Agnostic Text Embeddings](https://mlsanity.com/paper/2507.07125/) (2025)
- [Hard-aware Instance Adaptive Self-training for Unsupervised Cross-domain Semantic Segmentation](https://mlsanity.com/paper/2302.06992/) (2023)
- [Subjective Camera 1.0: Bridging Human Cognition and Visual Reconstruction through Sequence-Aware Sketch-Guided Diffusion](https://mlsanity.com/paper/2506.23711/) (2025)
- [HardBoost: Boosting Zero-Shot Learning with Hard Classes](https://mlsanity.com/paper/2201.05479/) (2022)
- [Reinforcement Learning for Unsupervised Domain Adaptation in Spatio-Temporal Echocardiography Segmentation](https://mlsanity.com/paper/2510.14244/) (2025)

## Built on by (most cited first, 1 of 1)

- [Ultra: Unsupervised Cross-Task Optimization for Reliable Restoration Segmentation Collaboration under Adverse Weather](https://mlsanity.com/paper/2608.16589/) (2026)

## Abstract (arXiv)

The learning order of semantic classes significantly impacts unsupervised domain adaptation for semantic segmentation, especially under adverse weather conditions. Most existing curricula rely on handcrafted heuristics (e.g., fixed uncertainty metrics) and follow a static schedule, which fails to adapt to a model's evolving, high-dimensional training dynamics, leading to category bias. Inspired by Reinforcement Learning, we cast curriculum learning as a sequential decision problem and propose an autonomous class scheduler. This scheduler consists of two components: (i) a high-dimensional state encoder that maps the model's training status into a latent space and distills key features indicative of progress, and (ii) a category-fair policy-gradient objective that ensures balanced improvement across classes. Coupled with mixed source-target supervision, the learned class rankings direct the network's focus to the most informative classes at each stage, enabling more adaptive and dynamic learning. It is worth noting that our method achieves state-of-the-art performance on three widely used benchmarks (e.g., ACDC, Dark Zurich, and Nighttime Driving) and shows generalization ability in synthetic-to-real semantic segmentation.