Kajian Komparatif Model 3D Berbasis Multi-View Stereo dan 3D Gaussian Splatting (Studi Kasus: Fakultas Teknik, UGM)

Ilham Nanda Ryantama, Ir. Ruli Andaru, S.T., M.Eng., Ph.D.

2026 | Skripsi | TEKNIK GEODESI

Pemodelan tiga dimensi (3D) untuk kawasan urban membutuhkan format representasi spasial yang tidak hanya unggul dari aspek fotorealistik, melainkan juga akurat secara geometri dan efisien dalam pengelolaan ruang penyimpanannya. Pendekatan rekonstruksi berbasis algoritma Multi-View Stereo (MVS) membutuhkan waktu pemrosesan yang lama pada tahapan pembentukan awan titik padat (dense point cloud) serta menghasilkan ukuran berkas yang besar. Selain itu, representasi hasil rekonstruksi algoritma MVS memiliki struktur entitas diskrit yang menyisakan ruang kosong, sehingga visualisasi objek akan terlihat retak atau berlubang saat diamati dari jarak dekat (close-range view). Sebagai alternatif solusi untuk mengatasi keterbatasan tersebut, algoritma 3D Gaussian Splatting (3DGS) hadir menawarkan pendekatan representasi berbasis partikel volumetrik dengan kemampuan perenderan secara real-time. Penelitian ini bertujuan untuk melakukan kajian komparatif performa rekonstruksi model 3D antara algoritma MVS dan algoritma 3DGS guna mengevaluasi kualitas produk spasial berskala kawasan.  Pemetaan dilakukan pada kawasan FT UGM memanfaatkan data foto udara low oblique ber-geotagging Post-Processing Kinematic (PPK). Alur pemrosesan data dilakukan secara independen pada kedua metode melalui pemanfaatan perangkat lunak pemrosesan fotogrametri untuk algoritma MVS dan framework pelatihan citra untuk algoritma 3DGS. Setelah model 3D berhasil terbentuk secara independen dari kedua algoritma pemrosesan tersebut, tahapan penyelarasan referensi spasial relatif dijalankan menggunakan perangkat lunak CloudCompare melalui kombinasi metode Point-Pair Picking (PPP) dengan skala metrik terkunci 1,0 serta optimasi Iterative Closest Point (ICP) yang bekerja meminimalkan jarak antartitik secara iteratif. Evaluasi multi-aspek kemudian dilakukan melalui uji kerapatan permukaan (surface density), analisis deviasi ortogonal spasial kuantitatif global dan lokal menerapkan algoritma Cloud-to-Mesh (C2M) Distance, serta uji signifikansi terhadap konsistensi dimensi metrik objek fisik, di samping membandingkan efisiensi perenderan (rendering) serta kualitas visualnya.  Hasil penelitian menunjukkan algoritma MVS menghasilkan struktur titik diskrit masif dengan kerapatan 744.686 pts/m2 berukuran berkas 5,75 GB, sementara 3DGS beroperasi jauh lebih efisien dengan kerapatan pusat partikel sebesar 42.208 pts/m2 berukuran berkas 730 MB. Reduksi geometri ini memangkas konsumsi memori grafis dari 19 GB VRAM menjadi 5,26 GB VRAM serta sukses mengeliminasi anomali topologi (melting artifacts) pada material kaca melalui implementasi fungsi Spherical Harmonics (SH). Pasca-registrasi spasial relatif, uji global C2M menghasilkan nilai mean distance sebesar 0,401243 cm dengan standar deviasi 40,1297 cm akibat outliers pada objek non-statis. Pengujian lokal pada Area of Interest (AOI) objek solid menunjukkan tingkat keselarasan tinggi dengan nilai mean distance sebesar 0,1490 cm pada segmen atap, 5,7452 cm pada dinding fasad, dan 12,4974 cm pada paving jalan. Dengan keunggulan kontinuitas geometri permukaan elipsoid (splats) dibandingkan dense cloud, serta kemampuannya menjaga kedetailan rekonstruksi 3D secara fotorealistik, model 3DGS terbukti memberikan performa spasial paling optimal sekaligus konsisten

Three-dimensional (3D) modeling of heterogeneous urban areas requires spatial representation formats that are not only photorealistic but also geometrically accurate and storage-efficient. Reconstruction approaches based on the Multi-View Stereo (MVS) algorithm demand extensive processing time during the dense point cloud generation phase and inherently produce large file sizes. Furthermore, MVS-derived reconstructions exhibit a discrete entity structure that leaves hollow gaps, causing object visualizations to appear fractured or fragmented from a close-range view. To overcome these limitations, the 3D Gaussian Splatting (3DGS) algorithm emerges as an alternative solution, offering a volumetric particle-based representation equipped with real-time rendering capabilities. Consequently, this study aims to perform a comparative analysis of 3D model reconstruction performance between the MVS and 3DGS algorithms to evaluate the quality of spatial products on a regional scale.  Spatial mapping was conducted across the FT UGM region utilizing low oblique aerial photography embedded with Post-Processing Kinematic (PPK) geotagging. The data processing workflows were independently executed for both methods, employing photogrammetric processing software for the MVS algorithm and an image training framework for the 3DGS algorithm. Upon extracting the geometric structures from each respective system, relative spatial reference alignment was performed within CloudCompare software through a combination of manual Point-Pair Picking (PPP) with a locked metric scale factor of 1.0 and Iterative Closest Point (ICP) optimization that works by iteratively minimizing the distance between points. A multi-aspect evaluation was subsequently carried out, encompassing surface density tests, macro-global and micro-local quantitative spatial orthogonal deviation analyses utilizing the Cloud-to-Mesh (C2M) Distance algorithm, and significance testing regarding the dimensional consistency of physical object metrics, alongside comparing rendering efficiency and visual quality.  The results indicate that the MVS algorithm generated a massive discrete point structure with a surface density of 744.686 pts/m2 and a file size of 5.75 GB, whereas 3DGS operated significantly more efficiently with a point center density of 42.208 pts/m2 and a compressed file size of 730 MB. This geometric reduction minimized graphic hardware load from 19 GB of VRAM to 5.26 GB of VRAM, while successfully eliminating melting artifacts on reflective glass surfaces through the implementation of the Spherical Harmonics (SH) function. Post-spatial registration, the macro-global C2M evaluation yielded a mean distance of 0.401243 cm with a standard deviation of 40.1297 cm, primarily induced by outlier distribution across dynamic non-rigid objects. Micro-local testing on specific Area of Interest (AOI) segments demonstrated high geometric stability with mean distances of 0.1490 cm for the roof segment, 5.7452 cm for the vertical glass facade, and 12.4974 cm for the roadway pavement. Given the superior surface continuity of ellipsoid geometry (splats) compared to dense clouds, along with its capability to preserve photorealistic 3D reconstruction details, the 3DGS model is proven to deliver the most optimal and highly consistent spatial performance.  

Kata Kunci : Fotogrametri, MVS, 3D Gaussian Splatting, Cloud-to-Mesh, Registrasi Spasial