| Proj No. | A1050-251 |
| Title | The Anatomy of YOLO: An Investigation into Optimisation for Mobile Devices |
| Summary | The real-time object detection capabilities of YOLO (You Only Look Once) have made it widely adopted in a multitude of domains. While its performance on desktop-class hardware is well-documented, its performance on mobile devices, which have lower computational resources, is comparatively lacking. Furthermore, there are limited research in examining how the architecture design choices of YOLO translate to actual performance on mobile hardware. This lack of concrete guidelines on appropriate model selection necessitates extensive empirical trial-and-error. This project attempts to address the identified gaps by investigating the relationship between YOLO's architecture and its performance on mobile hardware. Focusing on YOLOv11 as a testbed, controlled ablation experiments will be conducted on select architectural components of the model. Models will be trained and quantized before being deployed on a Samsung Galaxy S24 FE, where key metrics including, but not limited to FPS (frames-per-second), accuracy and memory usage will be collected. The findings aim to clarify how the architectural modifications affect efficiency-accuracy trade-offs, providing insights to guide optimised YOLO deployment on edge devices. |
| Supervisor | A/P Ling Keck Voon (Loc:S2 > S2 B2A > S2 B2A 22, Ext: +65 67905567) |
| Co-Supervisor | - |
| RI Co-Supervisor | - |
| Lab | Computer Engineering II (Loc: S2-B3b-08) |
| Single/Group: | Single |
| Area: | Digital Media Processing and Computer Engineering |
| ISP/RI/SMP/SCP?: |