Pima Hani Safitri
Program Studi Informatika, Universitas Telkom, Kampus Surabaya, Indonesia

Published : 1 Documents Claim Missing Document
Claim Missing Document
Check
Articles

Found 1 Documents
Search

RC-MH-GRU untuk Prediksi Arah Pandang 360° pada Telepresensi Robot Abdullah Iskandar; Pima Hani Safitri; Jati Hiliamsyah Husen; Muhammad Arifin
BIOS : Jurnal Teknologi Informasi dan Rekayasa Komputer Vol 8 No 1 (2027): March (In Progress)
Publisher : Puslitbang Sinergis Asa Professional

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.37148/bios.v8i1.249

Abstract

Viewport prediction can help a 360-degree video system prioritize panoramic regions for tile-based, region-of-interest (ROI), or adaptive-quality streaming. This study evaluates a Ridge-calibrated multi-horizon gated recurrent unit (RC-MH-GRU) offline at 100, 200, 300, and 500 ms. The aggregate dataset contains 72 videos and 2,986 head-tracking sessions sampled at 10 Hz. Each two-degree-of-freedom viewing direction is mapped to a three-dimensional unit vector; one second of position and first-difference history forms the input. Ridge-AR and the 14,604-parameter MH-GRU are combined on the unit sphere with one validation-selected weight per horizon, yielding 15,336 parameters. Evaluation covers five training seeds, a video-disjoint test split, five-fold video cross-validation, strong baselines, ablation, computational profiling, video-level statistics, and a parameter-matched 14,412-parameter Tiny Transformer. Across 60,000 test windows, mean angular errors are 0.958°, 2.033°, 3.324°, and 6.041°, lower than Ridge-AR at all horizons and compared with 1.026°, 2.090°, 3.383°, and 6.118° for the Tiny Transformer. Cross-validation on 72 videos also favors RC-MH-GRU over Ridge-AR at 200–500 ms. The evidence supports a compact prediction component for prospective tile/ROI allocation, not bandwidth savings, viewport-quality gains, or end-to-end telepresence performance.