Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SS-SFDA : Self-Supervised Source-Free Domain Adaptation for Road Segmentation in Hazardous Environments

Divya Kothandaraman, Rohan Chandra, Dinesh Manocha

TL;DR

The paper tackles road segmentation under adverse weather using a novel Self-Supervised Source-Free Domain Adaptation (SS-SFDA) framework. It pre-trains a self-attention-based auto-encoder on clear weather data and then adapts to unlabeled target domains via a two-step, self-supervised process: entropy minimization to enrich pseudo-labels, followed by online self-training with curriculum learning; a few-image distillation stage further refines performance on heterogeneous weather data. Across six real and synthetic datasets, SS-SFDA achieves 88–96% of supervised mIoU, outperforms prior SFDA methods by up to 10.26% on real adverse-weather data, and delivers 18–180× faster training than GAN-based SFDA. The approach advances robust autonomous driving perception in hazardous conditions, while acknowledging limitations in multi-class segmentation and suggesting extensions to other CV tasks in future work.

Abstract

We present a novel approach for unsupervised road segmentation in adverse weather conditions such as rain or fog. This includes a new algorithm for source-free domain adaptation (SFDA) using self-supervised learning. Moreover, our approach uses several techniques to address various challenges in SFDA and improve performance, including online generation of pseudo-labels and self-attention as well as use of curriculum learning, entropy minimization and model distillation. We have evaluated the performance on $6$ datasets corresponding to real and synthetic adverse weather conditions. Our method outperforms all prior works on unsupervised road segmentation and SFDA by at least 10.26%, and improves the training time by 18-180x. Moreover, our self-supervised algorithm exhibits similar accuracy performance in terms of mIOU score as compared to prior supervised methods.

SS-SFDA : Self-Supervised Source-Free Domain Adaptation for Road Segmentation in Hazardous Environments

TL;DR

The paper tackles road segmentation under adverse weather using a novel Self-Supervised Source-Free Domain Adaptation (SS-SFDA) framework. It pre-trains a self-attention-based auto-encoder on clear weather data and then adapts to unlabeled target domains via a two-step, self-supervised process: entropy minimization to enrich pseudo-labels, followed by online self-training with curriculum learning; a few-image distillation stage further refines performance on heterogeneous weather data. Across six real and synthetic datasets, SS-SFDA achieves 88–96% of supervised mIoU, outperforms prior SFDA methods by up to 10.26% on real adverse-weather data, and delivers 18–180× faster training than GAN-based SFDA. The approach advances robust autonomous driving perception in hazardous conditions, while acknowledging limitations in multi-class segmentation and suggesting extensions to other CV tasks in future work.

Abstract

We present a novel approach for unsupervised road segmentation in adverse weather conditions such as rain or fog. This includes a new algorithm for source-free domain adaptation (SFDA) using self-supervised learning. Moreover, our approach uses several techniques to address various challenges in SFDA and improve performance, including online generation of pseudo-labels and self-attention as well as use of curriculum learning, entropy minimization and model distillation. We have evaluated the performance on datasets corresponding to real and synthetic adverse weather conditions. Our method outperforms all prior works on unsupervised road segmentation and SFDA by at least 10.26%, and improves the training time by 18-180x. Moreover, our self-supervised algorithm exhibits similar accuracy performance in terms of mIOU score as compared to prior supervised methods.

Paper Structure

This paper contains 20 sections, 5 equations, 3 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Main Contributions
  3. Related Work
  4. Road Segmentation
  5. Domain Adaptation and Source Free Domain Adaptation
  6. Self-Supervised Learning
  7. Our Approach
  8. Pre-Training Baselines Using Self-Attention
  9. SS-SFDA
  10. Step 1: Bridging the Domain Gap via Entropy Minimization
  11. Step 2: Online Self-Training Using Enriched Pseudo-Labels
  12. Few-Image Fine-Tuning via Model Distillation
  13. Experiments and Results
  14. Analysing Pre-Trained Self-Attention-based AutoEncoder
  15. Results on Synthetic Datasets: Rain and Fog
  16. ...and 5 more sections

Figures (3)

  • Figure 1: We highlight the results generated by SS-SFDA on night sakaridis2019guided and fog benchmarks sakaridis2018model, compared to the baseline source model pre-trained on clear weather CityScapes. The purple regions (right) denote the segmented road pixels. The overall accuracy of our self-supervised algorithm in terms of mIoU is ($88-96\%$) of supervised methods.
  • Figure 2: Our Approach: In stage $1$, our model is pre-trained on a clear weather source dataset. In stage $2$, our model is initialized with the pre-trained model from stage $1$ and trained using our self-supervised algorithm, SS-SFDA, on the unlabeled adverse weather dataset. For heterogeneous weather datasets, we perform additional refinement steps based on model distillation (stage $3$).
  • Figure 3: Qualitative results. Our model generates results that closely resemble the ground-truth (GT) compared to the baseline CityScapes pre-trained model. Purple indicates the segmented road region. More results can be found in the supplementary material.