Engineers have developed an artificial intelligence framework that recognizes hand-drawn automotive styling sketches to retrieve and recommend matching three-dimensional digital vehicle models, significantly accelerating vehicle body development. In styling design, translating freehand two-dimensional concept drawings into engineering-ready three-dimensional surfaces has traditionally represented an inefficient bottleneck where original aesthetic intent is easily lost. To address this transition, researchers designed an end-to-end deep learning architecture called the Cross-domain Attention Multi-view Fusion network, designated CAMTF. The system maps the subjective, artistic contours of hand-drawn vehicle sketches directly into a shared mathematical embedding space alongside standardized multi-view projections of historical vehicle models. By matching conceptual line features with geometric CAD surfaces, the methodology establishes a digital pipeline for historical model reuse, allowing designers to bypass from-scratch geometric surface construction during initial concept phases.
Concept styling in vehicle manufacturing relies heavily on freehand sketching characterized by fluid lines, exaggerated proportions, and non-fixed perspective viewpoints. While traditional computer-aided design and mechanical feature recognition tools depend on rigid geometric constraints and predefined rule libraries suited for prismatic parts like shafts or boxes, automotive free-form exterior surfaces like roof panels and doors present complex topological curves that evade conventional rule-based algorithms. Hand-drawn sketches inevitably exhibit visual ambiguities such as uneven line thickness, overlapping strokes, scale inconsistencies, and background noise. Furthermore, general-purpose sketch-based three-dimensional retrieval models trained on standard everyday object datasets struggle to capture the domain-specific proportions, structural layouts, and subtle styling semantics unique to road vehicles, frequently suffering from severe feature distribution shifts across modalities.
To overcome sketch ambiguities, the framework introduces a multi-operator fusion preprocessing pipeline that sharpens styling lines while suppressing non-structural background noise. The input sketch first undergoes smoothing across three distinct scale spaces using a multi-scale Laplacian of Gaussian operator, with smoothing parameters and a scaling weight coefficient applied to emphasize both fine and broad contour lines. To capture directional characteristics, the system applies a direction-aware Scharr operator that computes first-order horizontal and vertical gradients, generating an edge magnitude map that is subsequently smoothed through a five-by-five mean filter. This smoothed gradient map acts as a structural saliency mask, multiplied element-wise by the Laplacian sharpened image to confine detail enhancement strictly to prominent edge boundaries. After linearly recombining this masked output with the original sketch using a fusion coefficient of 0.5, a final Gamma correction parameter of 0.8 adjusts grayscale distribution, yielding stable, noise-free vehicle contours.
Complementing sketch preprocessing, candidate three-dimensional CAD models are systematically converted into standardized two-dimensional multi-view projections to establish stable visual representations. Because styling sketches typically emphasize isometric, oblique axonometric, and side-profile perspectives, the researchers established a twelve-view spherical projection method. Using the vertical Z-axis of the normalized vehicle model as the rotation axis, viewpoints are elevated at a fixed elevation angle of 30 degrees and captured at uniform 30-degree intervals across a full circle. This fixed-viewpoint projection yields twelve equally spaced views covering all major exterior orientations, ensuring that key styling elements along the vehicle front, rear, fenders, and side silhouettes are captured in standardized inputs suitable for multi-view feature extraction.
At the architectural core, feature extraction and cross-domain alignment are driven by deep residual networks equipped with channel attention. Both the preprocessed two-dimensional sketches and the twelve projection views are fed into a shared backbone based on SE-ResNet-50 or SE-ResNet-152 architectures. The residual blocks employ combinations of one-by-one convolutions for channel compression and three-by-three convolutions to systematically expand receptive fields, transitioning from local line segments to vehicle-level proportional geometries. Squeeze-and-Excitation attention modules then perform global average pooling followed by compression-activation-expansion transformations with a compression ratio of 16 to dynamically weight semantic channels. The resulting sketch feature vector serves as a query vector, while the twelve extracted multi-view feature vectors serve as key and value representations within a stacked two-layer cross-domain attention mechanism featuring eight attention heads.
To train the cross-domain representation effectively, the researchers implemented an Adaptive Multi-level Contrastive Metric Learning strategy, termed AMCML. This joint optimization strategy balances three distinct loss components via explicit hyperparameters: an InfoNCE loss, a triplet loss, and an attention entropy regularization term. The InfoNCE global contrastive loss enforces high-level semantic alignment by pulling matching sketch and three-dimensional model pairs together while repelling non-matching negative samples across training batches. The triplet loss introduces an explicit safety margin buffer to sharpen sensitivity toward local geometric variations, ensuring that positive sample distances remain distinctly smaller than negative sample distances. Concurrently, the attention entropy regularization term penalizes dispersed attention distributions across the projection views, actively encouraging the multi-head attention blocks to concentrate weight on the specific projection angles that match the sketch viewpoint.
To validate the architecture, the researchers assembled a specialized automotive dataset consisting of 635 paired sketch and three-dimensional model instances spanning 15 distinct vehicle categories, encompassing sedans, sports cars, and sport utility vehicles. Paired models were divided into an 80 percent training subset and a 20 percent testing subset. In comprehensive retrieval evaluations on a single NVIDIA RTX A6000 graphics processing unit, the proposed CAMTF network achieved superior performance across standard cross-modal metrics. The best-performing 152-layer configuration achieved a Nearest Neighbor accuracy of 0.879, a First Tier recall of 0.798, a Second Tier recall of 0.845, an E-measure of 0.904, a normalized Discounted Cumulative Gain of 0.921, and a mean Average Precision of 0.856. In comparative testing against eleven benchmark models, CAMTF surpassed the leading prior method, Sequential Learning with Cross-Modal Representation, which had attained a mean Average Precision of 0.822, representing an approximate 4.1 percent relative performance improvement.
Ablation experiments systematically confirmed the individual contributions of each architectural module. Removing the cross-domain attention fusion mechanism and directly concatenating features caused a steep drop across all metrics, while utilizing only a single attention layer proved insufficient to resolve fine-grained correspondences between sketch lines and three-dimensional surfaces. Similarly, isolating the loss functions demonstrated that single-objective training could not maintain both intra-class compactness and inter-class separation, whereas combining all three objectives in AMCML produced the most discriminative feature space. Viewpoint ablation further revealed that decreasing projection views from twelve down to six or one severely degraded retrieval quality due to missing geometric perspectives, and replacing the CAMTF attention fusion with standard average pooling produced consistently inferior matching across all view configurations.
Beyond computational metric benchmarking, the researchers conducted engineering-oriented interactive modeling experiments to evaluate real-world workflow impacts. Seven professional automotive surface designers were invited to model vehicle exterior surfaces corresponding to five distinct styling sketch test cases under three comparative methodologies: traditional modeling from scratch, manual experience-based historical model reuse, and the proposed CAMTF retrieval-based reuse framework. Tasks were concluded once the resulting digital surface fully satisfied the design intent established by the original sketch. For the primary evaluation sketch, modeling times were tracked across all seven surface engineers, and average design times across all five vehicle cases were systematically compiled.
The engineering experiments revealed that the retrieval-based assisted design method reduced the average exterior modeling time across the five styling cases to 30.3 hours. In comparison, traditional from-scratch modeling required an average of 150.8 hours, representing an 80.0 percent reduction in modeling duration, while manual experience-based reuse required 39.9 hours, indicating an approximate 24.1 percent time reduction achieved by CAMTF. The authors noted that while both reuse strategies bypass blank-canvas modeling, human stylists struggle to manually search through large legacy databases to locate the exact three-dimensional model that best embodies complex sketch semantics. By automatically scoring design intent similarity via Euclidean distance in the learned embedding space, the network recommended optimal historical reference geometries requiring only local boundary adjustments and surface fairing.
Despite these measurable efficiency gains, the researchers highlighted several technical boundaries and engineering limitations inherent to the current study. The constructed paired dataset of 635 instances remains modest in scale and does not capture the full range of personal drawing styles, commercial branding trends, or stages of conceptual development across the global automotive industry. Generalization across diverse external datasets, unique styling aesthetics, and alternative vehicle classes will require further validation. Additionally, the framework operates as an assisted retrieval-and-reuse recommendation system rather than an automated generative model; human stylists must still perform manual surface fairing, Class-A reconstruction, and CAD adjustments on the retrieved geometry. Finally, the current network accepts only a single two-dimensional sketch query at a time, leaving the multi-sketch fusion of simultaneous front, side, and isometric drawings as an avenue for future development.
Subject of Research: Sketch-based cross-domain retrieval and assisted three-dimensional automotive styling design
Article Title: An automotive 3D modeling-assisted design method based on 2D sketch feature recognition
Article References: Sheng, G., Lyu, J., Sun, L., Guo, R., & Xiong, D. (2026). An automotive 3D modeling-assisted design method based on 2D sketch feature recognition. Journal of King Saud University Computer and Information Sciences, 38(9), Article 994. https://doi.org/10.1007/s44443-026-01454-3
Image Credits: AI Generated
DOI: 10.1007/s44443-026-01454-3
Keywords: sketch-based 3D model retrieval, automotive styling design, cross-domain attention, multi-view fusion, computer-aided styling, SE-ResNet, deep metric learning, contrastive learning, Laplacian of Gaussian, Scharr operator, surface modeling, feature recognition, digital vehicle design, Transformer attention, CAD model reuse, Euclidean distance matching, engineering design automation
Source: bioengineer.org



