Estimasi Volumetrik Objek Dari Kamera Rgb Menggunakan Yolo Instance Segmentation Dan Monocular Depth Estimation

Afrina, Rumaisha (2026) Estimasi Volumetrik Objek Dari Kamera Rgb Menggunakan Yolo Instance Segmentation Dan Monocular Depth Estimation. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221146-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (9MB) | Request a copy

Abstract

Pemantauan sampah perairan dan budi daya rumput laut memerlukan estimasi volume objek yang cepat dan terjangkau, namun metode konvensional umumnya bergantung pada sensor mahal atau pengukuran manual. Penelitian ini membangun pipeline estimasi volumetrik objek berbasis kamera RGB tunggal yang memadukan YOLO instance segmentation dan Monocular Depth Estimation (MDE), serta diuji pada dua studi kasus yaitu, sampah terapung dengan metode convex hull dan sampel rumput laut dengan metode Panjang × Lebar × Tinggi (PLT). Empat varian YOLO dan enam model MDE dibandingkan secara empiris terhadap volume ground truth. Hasil menunjukkan YOLOv8s-seg terbaik pada dataset sampah dengan nilai mask mAP@50–95 sebesar 0,926 dan YOLOv11s-seg terbaik pada dataset rumput laut dengan nilai mask mAP@50-95 sebesar 0,8786. Untuk estimasi volume, ZoeDepth-KITTI paling akurat pada sampah dengan akurasi 97,08%, sedangkan Depth-Anything V2 Large paling akurat pada rumput laut dengan akurasi 96,95%. Penelitian ini menyimpulkan bahwa pemilihan model MDE bergantung pada metode volume yang digunakan. Model kedalaman metrik unggul untuk rekonstruksi convex hull, sedangkan model kedalaman relatif unggul untuk metode PLT. Pipeline yang dibangun terbukti akurat untuk estimasi volume objek perairan.
=======================================================================================================================================
Aquatic litter monitoring and seaweed cultivation require fast and affordable object volume estimation, yet conventional methods generally rely on expensive sensors or manual measurement. This study develops a single RGB camera-based volumetric object estimation pipeline that integrates YOLO instance segmentation and Monocular Depth Estimation (MDE), evaluated on two case studies which are floating waste using the convex hull method and seaweed samples using the Length × Width × Height (LWH) method. Four YOLO variants and six MDE models were empirically compared against ground truth volumes. The results show that YOLOv8s-seg performed best on the waste dataset with the mask mAP@50–95 score of 0.926 and YOLOv11s-seg performed best on the seaweed dataset with the mask mAP@50-95 score of 0.8786. For volume estimation, ZoeDepth-KITTI was the most accurate on waste with 97.08% accuracy, whereas Depth-Anything V2 Large was the most accurate on seaweed with 96.95% accuracy. This study concludes that MDE model selection depends on the volume method used. Metric depth models excel for convex hull reconstruction, while relative depth models excel for the LWH method. The proposed pipeline proves to be accurate for estimating the volume of aquatic objects.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Convex Hull, Estimasi Volume, Instance Segmentation, Monocular Depth Estimation, YOLO, Volume Estimation
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.P3 Pattern recognition systems
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Rumaisha Afrina
Date Deposited:	20 Jul 2026 04:50
Last Modified:	20 Jul 2026 06:14
URI:	http://repository.its.ac.id/id/eprint/135599

Actions (login required)

View Item