Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Technical Report for ICRA 2025 GOOSE 2D Semantic Segmentation Challenge: Boosting Off-Road Segmentation via Photometric Distortion and Exponential Moving Average

Wonjune Kim, Lae-kyoung Lee, Su-Yong An

TL;DR

This work targets robust off-road semantic segmentation with extreme appearance variability and class imbalance. It leverages a FlashInternImage-B backbone and a UPerNet decoder, enhanced by photometric distortion augmentation and an Exponential Moving Average of weights to improve generalization. Ablation shows photometric jitter contributes +0.48 mIoU and EMA contributes +1.12 mIoU, achieving 88.8 mIoU on the GOOSE/GOOSE-EX validation, with strong gains for underrepresented classes like obstacle and human. On the official test, the approach reaches 84.5 mIoU and ranks second on the public leaderboard, underscoring the practicality of combining established components with targeted augmentations for off-road perception.

Abstract

We report on the application of a high-capacity semantic segmentation pipeline to the GOOSE 2D Semantic Segmentation Challenge for unstructured off-road environments. Using a FlashInternImage-B backbone together with a UPerNet decoder, we adapt established techniques, rather than designing new ones, to the distinctive conditions of off-road scenes. Our training recipe couples strong photometric distortion augmentation (to emulate the wide lighting variations of outdoor terrain) with an Exponential Moving Average (EMA) of weights for better generalization. Using only the GOOSE training dataset, we achieve 88.8\% mIoU on the validation set.

Technical Report for ICRA 2025 GOOSE 2D Semantic Segmentation Challenge: Boosting Off-Road Segmentation via Photometric Distortion and Exponential Moving Average

TL;DR

This work targets robust off-road semantic segmentation with extreme appearance variability and class imbalance. It leverages a FlashInternImage-B backbone and a UPerNet decoder, enhanced by photometric distortion augmentation and an Exponential Moving Average of weights to improve generalization. Ablation shows photometric jitter contributes +0.48 mIoU and EMA contributes +1.12 mIoU, achieving 88.8 mIoU on the GOOSE/GOOSE-EX validation, with strong gains for underrepresented classes like obstacle and human. On the official test, the approach reaches 84.5 mIoU and ranks second on the public leaderboard, underscoring the practicality of combining established components with targeted augmentations for off-road perception.

Abstract

We report on the application of a high-capacity semantic segmentation pipeline to the GOOSE 2D Semantic Segmentation Challenge for unstructured off-road environments. Using a FlashInternImage-B backbone together with a UPerNet decoder, we adapt established techniques, rather than designing new ones, to the distinctive conditions of off-road scenes. Our training recipe couples strong photometric distortion augmentation (to emulate the wide lighting variations of outdoor terrain) with an Exponential Moving Average (EMA) of weights for better generalization. Using only the GOOSE training dataset, we achieve 88.8\% mIoU on the validation set.
Paper Structure (10 sections, 1 equation, 2 figures, 1 table)

This paper contains 10 sections, 1 equation, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Method
  3. Baseline
  4. Photometric Distortion
  5. Exponential Moving Average
  6. Experiments
  7. Dataset and Training Protocol
  8. Quantitative Results
  9. Qualitative Results
  10. Conclusions

Figures (2)

  • Figure 1: Qualitative comparison on the validation set. Columns from left to right: (1) input RGB image, (2) prediction of the FlashInternImage-B baseline, (3) baseline plus photometric distortion, (4) baseline plus photometric distortion and EMA, and (5) ground-truth annotation.
  • Figure 2: Comprehensive Photometric Distortion Effects. From left to right and top to bottom the grid shows (i) the original RGB image, (ii) a combined (+) sample where brightness, contrast, saturation and hue are jointly increased, (iii) a combined (–) sample where the same factors are jointly decreased, followed by isolated adjustments of brightness ($+40$ and $-40$), contrast ($\times1.3$ and $\times0.7$), and saturation ($\times1.3$ and $\times0.7$). These transformations are drawn at random during training, each with probability 0.5, to expose the network to the full range of illumination and color conditions encountered in off-road scenes.