Improving segmentation results of 3D meshes representing obese human body silhouettes with transfer learning technique

<p>The latest and most refined segmentation methods of 3D meshes representing human body silhouettes utilize a machine-learning approach. State-of-the-art models trained on broad datasets perform at a very high accuracy level and robustness, which is satisfactory for most applications in animation, medicine or the garment industry. However, for body shapes that vary from standard silhouette, results tend to be lower, especially for obese subjects with crotch, armpits and thigh areas that are difficult to analyse. The paper presents a process of preparing an additional dataset and labelling segments to increase the effectiveness of correct segmentation for these types of silhouettes. The authors provide a transfer learning technique for a convolutional neural network model benefiting from mapping 3D objects onto two-dimensional space that allows significant enhancement by over a 3-point percentage of mean accuracy. The outcomes surpass general methods, exceeding 96,56% of points correctly classified with a distinctively visible improvement in problematic areas.</p>

针对表征人体轮廓的三维网格（3D mesh）的最先进且最精细的分割方法，均采用机器学习范式。经大规模数据集训练的最先进模型，其分割精度与鲁棒性均处于极高水准，可满足动画、医学及服装行业的多数应用需求。然而，对于偏离标准轮廓的人体形态，分割效果往往欠佳；尤其是针对肥胖受试者的胯部、腋下与大腿区域——这些区域分析难度较高，分割效果更差。本研究提出了一套额外数据集构建与片段标注流程，旨在提升此类人体轮廓的正确分割效果。研究团队提出了一种面向卷积神经网络（Convolutional Neural Network）的迁移学习技术，该技术通过将三维物体映射至二维空间的思路，可使模型的平均准确率提升超过3个百分点，性能提升显著。该方法的分割效果优于通用基线方法，正确分类的网格点占比可达96.56%，且在此前的难点区域中，分割性能改善尤为突出。