The Impact of CutMix on Reliability and Robustness in Semantic Segmentation

Improving Robustness of Semantic Segmentation for Autonomous Driving: A Case Study

Deep neural networks (DNNs) have achieved remarkable success in recent years and are increasingly integrated into safety-critical systems such as autonomous driving vehicles. However, when deployed in real-world environments, their robustness to common input corruptions remains a major challenge for safety assurance. Corruptions such as motion blur can change the outputs of DNN-based semantic segmentation models and, more importantly, cause unsafe system-level decision inconsistencies, for example by failing to identify ground obstacles that are correctly recognized under clean conditions. In this paper, we present a testing-oriented robustness repair approach for semantic segmentation models in real-world industrial settings. We first use corruption-based testing to reveal decision-level failures under realistic perturbations, and then repair the model through a combination of data augmentation and self-training using only unlabeled data. Rather than focusing solely on pixel-level prediction changes, our method targets the reduction of system violations while preserving decision behavior on clean inputs. We evaluate the approach on a semantic segmentation model used in an industrial product. Experimental results show that our method significantly reduces system violation rates while maintaining system-level decision-making accuracy, demonstrating the practical value of testing-guided repair for safety-critical deployment.

Evaluating the Adaptation Potential of SAM2 for Glacier Segmentation in severe Weather

Experimental results demonstrate that the adapted SAM2 model achieves stable segmentation under moderate environmental variability, while degrading under severe visibility loss, consistent across model scales and input resolutions.

Comparative Analysis of Loss Functions for Semantic Segmentation: An Empirical Study on Cityscapes Dataset

— Semantic segmentation remains a fundamental challenge in computer vision, where the choice and weighting of loss functions significantly impact model performance. This study presents a comprehensive comparative analysis of individual versus combined loss functions with systematic weight ablation for semantic segmentation using modified Attention U-Net and DeepLabV3+ architectures on the Cityscapes dataset. We systematically evaluate seven weight configurations across three loss components (Cross-Entropy, Dice, Focal) through rigorous ablation studies, and validate our findings across two architectures to ensure generalizability. Through extensive experimentation across 20 epochs with 2,975 training and 500 validation images, our results demonstrate that the Dice-dominant weighting configuration (0.5:1.0:0.5 for CE:Dice:Focal) achieves superior performance with 57.83% mean Intersection over Union (mIoU) on Attention U-Net and 58.35% mIoU on DeepLabV3+, representing 7.78% improvement over the best individual loss function. Comprehensive ablation studies reveal that weight configuration critically affects performance, with Dice-dominant weighting consistently outperforming equal weighting (55.59% mIoU) and individual loss functions. Qualitative analysis demonstrates substantial improvements in boundary delineation and small object detection, with boundary IoU improving by 1.41% and challenging class performance (trucks, pedestrians) improving by 5 – 21%. Statistical analysis reveals that Cross-Entropy provides the most efficient training with a 75.4% loss reduction, while Dice loss exhibits convergence challenges, resulting in only a 34.5% reduction. Our findings conclusively demonstrate that optimized combined loss function weighting achieves better segmentation performance than both individual approaches and naive equal weighting strategies, with consistent improvements across different network architectures.

Computer Vision and Image Understanding

Semi-Supervised Structural Prior-Guided Network for Space Target Component Segmentation in ISAR Images

Highlights What are the main findings? The GMHC-ViT encoder broadens feature representation via multi-stream gating, effectively mitigating inter-class confusion, while the PGM injects shape and edge structural priors into weak-response channels, markedly improving boundary accuracy and component completeness. SSPNet consistently achieves superior segmentation performance across different annotation ratios and exhibits strong robustness under extremely limited labels and severe noise conditions. What are the implications of the main findings? The proposed SSPNet enables highly accurate and label-efficient space target component segmentation, greatly reducing the reliance on expensive pixel-level annotations in ISAR interpretation tasks. The work demonstrates the benefit of explicitly encoding domain-specific structural knowledge into deep networks, offering a valuable reference for related tasks. Abstract Segmenting key components of space targets using Inverse Synthetic Aperture Radar (ISAR) images is an important interpretation task in space situational awareness. However, the scarcity of pixel-level annotated data, inter-class confusion caused by morphological differences among multiple target classes, and the absence of structural priors for components restrict the performance improvement in existing deep models on this task. Therefore, this paper proposes a Semi-Supervised Structural Prior-Guided Network (SSPNet). First, a Gated Manifold-Constrained Hyper-Connections Vision Transformer (GMHC-ViT) encoder is proposed to broaden the feature representation space via parallel multi-feature streams with adaptive gating, thereby alleviating inter-class confusion and enhancing cross-category generalization. Second, a Prior-Guided Module (PGM) is proposed to extract shape and edge priors of components, and it adaptively enhances the weakly activated channels of encoder features through cross-attention, thereby injecting structural knowledge independent of image quality into the segmentation process. Furthermore, to effectively leverage large amounts of unlabeled data, a strong perturbation strategy tailored to the characteristics of ISAR images is designed for consistency regularization. Experimental results on a simulated ISAR dataset containing 38 classes of space targets demonstrate that SSPNet outperforms existing methods and exhibits strong segmentation capability even under low signal-to-noise ratio (SNR) conditions.

Automated Visual Inspection of Bridge Defect Segmentation Using Large-Scale Pretrained Models

Routine visual inspections of bridges are safety-critical activities that are still manual, time-consuming, and subjective to the inspector’s interpretation. This paper presents a scalable automated visual inspection pipeline for semantic segmentation of bridge defects, targeting the 19-class dacl10k benchmark dataset. We perform the first systematic comparison of pretraining paradigms on dacl10k, including CNN -supervised, ViT-supervised, masked image modelling, and self-supervised learning across 10 configurations under a controlled training protocol. The pretraining paradigm consistently dominates architectural choice, with DINOv2-L, pretrained on 142 million unlabelled images, achieving a mean Intersection-over-Union (mIoU) of 49.16%. Applying our native multi-label training approach to EVA-02-L, the dacl10k challenge-winning backbone, achieves 48.97% mIoU versus their 47.80% single-model result, demonstrating that training design is an independent performance factor. A three-model ensemble achieves 51.08% mIoU, exceeding the challenge winner’s score by using half the number of models. A prototype inspection system (InSpectralytiX) is deployed in a HuggingFace Gradio Space, demonstrating end-to-end feasibility from raw image to a per-class defect map. The future work targets automated condition scoring for bridge asset management integration, supporting structural health monitoring at the local level when performed repeatedly.