Abstract:

Objective The drape performance of fabrics constitutes one of their visual attributes, and investigating fabric drape performance holds significant reference value for both design optimization and retrieval applications. To date, fabric matching or search methodology based on drape models have primarily relied on extracting multiple drape-related metrics followed by comparative analysis among these quantified parameters. With the advancement of deep learning technologies, there arises an imperative to leverage deep learning frameworks for achieving automated fabric drape model matching. Consequently, this study proposes to conduct feature extraction and similarity computation for fabric drape models employing PointNet architecture, thereby establishing a systematic approach for precise drape-based fabric matching.
Method In order to extract feature vectors from fabric 3-D drape models automatically and to realize the matching of fabric 3-D drape models, 3-D drape models of 50 fabrics were collected using 3-D scanning devices. Based on the results of the fabric 3-D drape models, 11 fabrics were selected from the 50 fabrics, which follow obviously different 3-D drape models. The 3-D drape models of these 11 fabrics comprise a classification dataset D_PN, and the 3-D drape models of the remaining 39 fabrics are summarized into a test set D_RC, which is adopted to test the matching of the fabric 3-D drape models. Then the 3-D drape models in the datasets D_PN and D_RC are resampled to fabric 3-D drape models with the same number of vertices and topology using the resampling method. A PointNet classification model is built afterwards and the dataset D_PN is adopted to train the PointNet classification model. Finally, the trained PointNet is adopted to extract the feature vectors of the fabric 3-D drape model in the dataset D_RC, and the feature vectors are used as the basis to realize the matching of the fabric 3-D drape model. The recall of the dataset D_RC is adopted to evaluate the matching results of the fabric 3-D drape model. The influence of the number of vertices in the fabric 3-D drape model on the recall is also investigated.
Results The results show that the adopted PointNet can effectively realize the classification of the dataset D_PN. When the learning rate for training the PointNet model is fixed to 0.001, the PointNet model can be trained in 20 cycles and then the model can be seen to converge to a stable level. The loss function and classification accuracy reached a steady state for both the training and validation sets. The results also show that the feature vectors extracted using PointNet enable the matching of the fabric 3-D drape model. When PointNet is trained with uniformly sorted vertices, the maximum average recall of the dataset D_RC is 37.76%. When PointNet is trained with randomly ordered vertices and the number of vertices is 926, the maximum average recall of the dataset D_RC is 39.91% at the maximum value. This result is superior relative to the already reported manually designed fabric 3-D drape model feature IC_pca (38.56% recall).
Conclusion The PointNet model is employed to achieve feature extraction for fabric drape models and simultaneously realize fabric matching based on these models. The proposed method for extracting features from fabric drape models demonstrate greater convenience compared to conventional handcrafted design approaches. Furthermore, in the context of fabric matching using drape-based models, superior effectiveness is achieved relative to currently reported methods.

Key words: fabric, 3-D drape model, PointNet, deep learning algorithm, 3-D drape model matching, drape performance