Non-Stationary Class-Directed Augmentation for Fine-Grained Vehicle Damage Detection and Instance Segmentation

Automatedvehicle damage assessment remains challenging because damaged regions frequently exhibit irregular geometries, diffuse boundaries, substantial scale variability, and limited visual separability between categories such as cracks and scratches. Although previous studies have mainly improved detection and segmentation through architectural modifications, comparatively less attention has been paid to the augmentation regime used during training. This work investigates Non-Stationary Augmentation (NSA) for fine-grained vehicle damage analysis. In NSA, compositional transformations are applied with activation windows, selection probabilities, and internal parameters that vary over time, so that the augmented training distribution shifts progressively throughout optimization. A second variant, Non-Stationary Augmentation with Class-Directed sampling (NSA-CD), additionally incorporates class-directed source sampling. Both variants are evaluated on Car Damage Detection (CarDD) using Cascade Mask Region-based Convolutional Neural Network (R-CNN) with three ConvNeXt-V2 backbones of increasing capacity: Pico, Tiny, and Base. Relative to identical baseline configurations, both variants improve mask average precision (AP) by 1.2 to 4.3 points and box AP by 2.6 to 4.9 points across all three backbones, with the largest gains on small and medium objects, where box AP rises by 6.3 to 16.9 and 6.4 to 16.0 points, respectively. Whereas the baseline saturates early, the augmented configurations continue to improve until the end of training. Aggregate AP does not distinguish the two variants; at AP at an Intersection over Union (IoU) threshold of 0.50 (AP50), however, NSA-CD outperforms NSA on crack, the most difficult category, in five of six mask and box comparisons, with gains above 6 points for Pico and Base. In addition, both variants make Base the strongest configuration, surpassing Tiny; this ordering does not hold under baseline training. Each configuration was trained once, and the stability of each difference with respect to the test set was assessed through paired bootstrap resampling. These results suggest that a non-stationary compositional augmentation regime, optionally combined with class-directed sampling, is a practical complement to architectural improvements for fine-grained visual damage analysis.

Authors

Institutions

Publication Details

Journal
Applied Sciences
Published
2026-11-12
DOI
https://doi.org/10.3390/app162211239
Primary Topic
Advanced Neural Network Applications
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Non-Stationary Class-Directed Augmentation for Fine-Grained Vehicle Damage Detection and Instance Segmentation

Edison Atencio, Emanuel Vega, Rodrigo Fernando Herrera, Alejandro Cheuquepan
Applied Sciences
Advanced Neural Network Applications
article

Non-Stationary Class-Directed Augmentation for Fine-Grained Vehicle Damage Detection and Instance Segmentation

Edison Atencio, Emanuel Vega, Rodrigo Fernando Herrera, Alejandro Cheuquepan
article en

Abstract

Automatedvehicle damage assessment remains challenging because damaged regions frequently exhibit irregular geometries, diffuse boundaries, substantial scale variability, and limited visual separability between categories such as cracks and scratches. Although previous studies have mainly improved detection and segmentation through architectural modifications, comparatively less attention has been paid to the augmentation regime used during training. This work investigates Non-Stationary Augmentation (NSA) for fine-grained vehicle damage analysis. In NSA, compositional transformations are applied with activation windows, selection probabilities, and internal parameters that vary over time, so that the augmented training distribution shifts progressively throughout optimization. A second variant, Non-Stationary Augmentation with Class-Directed sampling (NSA-CD), additionally incorporates class-directed source sampling. Both variants are evaluated on Car Damage Detection (CarDD) using Cascade Mask Region-based Convolutional Neural Network (R-CNN) with three ConvNeXt-V2 backbones of increasing capacity: Pico, Tiny, and Base. Relative to identical baseline configurations, both variants improve mask average precision (AP) by 1.2 to 4.3 points and box AP by 2.6 to 4.9 points across all three backbones, with the largest gains on small and medium objects, where box AP rises by 6.3 to 16.9 and 6.4 to 16.0 points, respectively. Whereas the baseline saturates early, the augmented configurations continue to improve until the end of training. Aggregate AP does not distinguish the two variants; at AP at an Intersection over Union (IoU) threshold of 0.50 (AP50), however, NSA-CD outperforms NSA on crack, the most difficult category, in five of six mask and box comparisons, with gains above 6 points for Pico and Base. In addition, both variants make Base the strongest configuration, surpassing Tiny; this ordering does not hold under baseline training. Each configuration was trained once, and the stability of each difference with respect to the test set was assessed through paired bootstrap resampling. These results suggest that a non-stationary compositional augmentation regime, optionally combined with class-directed sampling, is a practical complement to architectural improvements for fine-grained visual damage analysis.

Applied SciencesVol. 16(22)
Pontificia Universidad Católica de Valparaíso (CL), Pontificial Catholic University of Valparaiso (CL)
Advanced Neural Network Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon