基于不确定相似性度量学习的三维模型草图检索

Uncertain Similarity Metric Learning for Sketch-Based 3D Model Retrieval

对于把手绘图像以及三维模型同时映射到一个联合特征嵌入空间和草图噪声给检索性能带来的影响,分析了跨模态匹配和草图噪声。将师生策略和数据不确定学习相结合,构建了一种基于不确定相似性度量学习的三维模型草图检索方法(uncertainty similarity metric learning, USML)。首先利用基于师生策略的三维形状语义相似度度量学习方法代替现有的跨域共享特征嵌入方法以提升检索的效率;然后,受数据不确定学习的启发,采用不确定学习来解决草图中噪声导致的严重过拟合并损害特征学习的问题。在大型公开基准数据集SHREC13和SHREC14上的实验结果验证了方法的有效性。结果表明,与深度点到子空间度量学习(depth point to subspace metric learning, DPSML)、深度相关度量学习(deep correlation metric learning, DCML)及深度相关整体度量学习(deep correlation holistic metric learning, DCHML)等未考虑噪声影响的算法相比,USML对基于抗噪草图的三维形状检索的效率提升至关重要。

Cross-modal matching and sketch noise were analyzed for the effect of mapping both hand-drawn images and 3D models into a joint feature embedding space and sketch noise on retrieval performance.Combining teacher-student strategy and data uncertainty learning,a 3D model sketch retrieval method based on uncertain similarity metric learning(USML)was constructed.The 3D shape semantic similarity metric learning method based on teacher-student strategy was used to replace the existing cross-domain shared feature embedding method to improve retrieval efficiency.Inspired by data uncertainty learning,uncertainty learning was employed to address the problem of severe overfitting caused by noise in sketches and impairing feature learning.Experimental results on large-scale public benchmark datasets SHREC13 and SHREC14 validated the effectiveness of the method.The results show that,compared with algorithms that do not consider the influence of noise,such as deep point-to-subspace metric learning(DPSML),deep correlated metric learning(DCML),and deep correlated holistic metric learning(DCHML),the USML is crucial for the efficiency improvement of noise-resistant sketch-based 3D shape retrieval.