Abiansyah, Birhasani Raka (2026) Optimisasi Mesh pada Model Single-view 3D Reconstruction Berbasis Deep Learning untuk Objek Topeng Bali. Other thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 5025221115-Undergraduate_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
5025221115-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (9MB) | Request a copy |
Abstract
Rekonstruksi mesh dari single-view 3D reconstruction merupakan salah satu tantangan dalam bidang visi komputer, terutama pada objek yang memiliki karakteristik bentuk tipis dan ornamen kompleks seperti topeng Bali. Perkembangan deep learning telah menghasilkan berbagai model image-to-3D modern, akan tetapi sebagian besar model tersebut dilatih menggunakan dataset objek umum sehingga performanya pada topeng Bali masih perlu diteliti. Penelitian ini bertujuan untuk menerapkan dan mengevaluasi model TripoSR, InstantMesh, dan Wonder3D dalam merekonstruksi mesh 3D dari single-view image pada dataset BaliMask3D. Penelitian ini juga mengimplementasikan fine-tuning pada TripoSR dan InstantMesh, serta menerapkan post-optimization pada mesh hasil rekonstruksi. Evaluasi dilakukan menggunakan metrik berbasis geometri, yaitu Chamfer Distance, Unidirectional Hausdorff Distance, F-score, dan Thickness Error, serta metrik berbasis render, yaitu Silhouette Intersection over Union, Depth MAE, dan Normal Cosine Similarity. Hasil penelitian menunjukkan bahwa Wonder3D menghasilkan visual baseline yang paling menyerupai topeng Bali, sedangkan fine-tuning meningkatkan kedekatan geometri pada TripoSR dan performa render pada InstantMesh. Setelah post-optimization, Wonder3D memperoleh hasil geometry-based terbaik dengan nilai CD 0.0719, UHD 0.0632, F-score 0.3746, dan Thickness Error 0.4563. Hasil ini menunjukkan bahwa optimisasi mesh dapat meningkatkan kualitas rekonstruksi 3D, terutama pada aspek kedekatan geometri dan proporsi ketebalan mesh.
======================================================================================================================================
Single-view 3D mesh reconstruction is one of the challenges in computer vision, particularly for objects with thin structures and complex ornaments, such as Balinese masks. The development of deep learning has introduced various modern image-to-3D models; however, most of these models are trained on general object datasets, and their performance on Balinese masks therefore requires further investigation. This study aims to implement and evaluate TripoSR, InstantMesh, and Wonder3D for reconstructing 3D meshes from single-view images using the BaliMask3D dataset. Fine-tuning is applied to TripoSR and InstantMesh, while post-optimization is performed on the reconstructed meshes. The evaluation uses geometry-based metrics, including Chamfer Distance, Unidirectional Hausdorff Distance, F-score, and Thickness Error, as well as render-based metrics, including Silhouette Intersection over Union, Depth MAE, and Normal Cosine Similarity. The results show that Wonder3D produces the baseline visualization that most closely resembles Balinese masks, while fine-tuning improves geometric similarity in TripoSR and render-based performance in InstantMesh. After post-optimization, Wonder3D achieves the best geometry-based performance, with a CD of 0.0719, UHD of 0.0632, F-score of 0.3746, and Thickness Error of 0.4563. These results demonstrate that mesh optimization can improve 3D reconstruction quality, particularly in terms of geometric similarity and mesh thickness proportions.
Actions (login required)
 |
View Item |