Jiefeng Lv

Jiefeng Lv


Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
6 November 2026 in Journal of King Saud University Computer and Information Sciences
Verified
Journal Article
Open access
PDF file
GSGanshu Sheng
Ganshu Sheng
JLJiefeng Lv
Jiefeng Lv
LSLuo Sun
Luo Sun
RGRuoyu Guo
Ruoyu Guo
DXDetian Xiong
Detian Xiong
To address the engineering challenges in automotive styling design, including the inefficient transition from two-dimensional (2D) sketches to three-dimensional (3D) surface modeling and the loss of design intent during this transition, an automotive 3D styling-assisted design assistance method based on a Cross-domain Attention Multi-view Fusion (CAMTF) network was proposed. First, a cascaded multi-operator strategy was employed to enhance sketch details and suppress noise in 2D sketches, while twelve-view rendering was used to obtain multi-view representations of 3D models. Subsequently, a Transformer-based cross-domain attention mechanism was introduced to unify the feature embedding space, and an adaptive multi-level contrastive metric learning strategy was incorporated to optimize the loss function, thereby achieving fine-grained semantic alignment of cross-domain features. Finally, the similarity between 2D sketches and candidate 3D models was quantified using the Euclidean distance between feature embeddings, through which design-intent similarity scores were obtained, and the most similar reference model was recommended for design assistance. Experimental results demonstrated that the proposed method achieved favorable retrieval performance on the automotive styling 2D sketch-3D model dataset constructed in this study. The best-performing configuration of the proposed network achieved NN, FT, ST, E-measure, nDCG, and mAP values of 0.879, 0.798, 0.845, 0.904, 0.921, and 0.856, respectively. Compared with SLCMR, which achieved the highest mAP among the existing methods evaluated in this study, the proposed method improved mAP from 0.822 to 0.856, corresponding to a relative improvement of approximately 4.1%. Furthermore, engineering-oriented assisted design experiments involving five automotive styling cases showed that the proposed retrieval-based reuse method reduced the average modeling time by approximately 24.1% compared with the manual experience-based reuse modeling method. These results indicate that, under the experimental conditions of this study, the proposed method can improve the efficiency of automotive styling model retrieval and provide effective support for 3D styling-assisted design based on the reuse of historical models.