Samsung R&D Institute Case Study
Samsung R&D Institute Case Study
Samsung R&D Institute
Samsung R&D Institute
As an AI Engineer Intern at Samsung R&D Institute India, I worked on single-view 3D reconstruction: turning one ordinary 2D photo into a usable 3D mesh. The work fed into on-device AI capabilities for Galaxy devices, so it had to be accurate, small and fast at once.
As an AI Engineer Intern at Samsung R&D Institute India, I worked on single-view 3D reconstruction: turning one ordinary 2D photo into a usable 3D mesh. The work fed into on-device AI capabilities for Galaxy devices, so it had to be accurate, small and fast at once.
2024
Applied AI Research
AI Engineer Intern, September 2024 to January 2025 · Samsung R&D Institute India · Bengaluru, India
Single-view 2D to 3D mesh reconstruction for on-device AI on Samsung Galaxy devices, built to hold up inside real phone latency and memory budgets.
Mission
Mission
I wanted the reconstruction work to survive contact with real hardware. Anything that only runs in a datacenter is a demo, so the goal from day one was a pipeline that produces a usable mesh from a single image and still fits inside a phone’s latency and memory budget.
I wanted the reconstruction work to survive contact with real hardware. Anything that only runs in a datacenter is a demo, so the goal from day one was a pipeline that produces a usable mesh from a single image and still fits inside a phone’s latency and memory budget.
Challenge
Challenge
Single-view reconstruction is underdetermined. One image gives you no ground truth for anything the camera couldn’t see, so the model has to infer geometry it has never observed. Getting that to research quality is hard enough; getting it inside the latency and memory limits of a Galaxy device is a second problem.
Single-view reconstruction is underdetermined. One image gives you no ground truth for anything the camera couldn’t see, so the model has to infer geometry it has never observed. Getting that to research quality is hard enough; getting it inside the latency and memory limits of a Galaxy device is a second problem.
Solution
Solution
I engineered a scalable pipeline for single-view 2D to 3D mesh reconstruction, then optimized inference directly against on-device latency and memory constraints instead of treating that as a port to do later.
I engineered a scalable pipeline for single-view 2D to 3D mesh reconstruction, then optimized inference directly against on-device latency and memory constraints instead of treating that as a port to do later.
Result
Result
The pipeline contributed to on-device AI capabilities for Samsung Galaxy devices, running inside the latency and memory budget that hardware allows, rather than staying a result that only holds up in a datacenter.
The pipeline contributed to on-device AI capabilities for Samsung Galaxy devices, running inside the latency and memory budget that hardware allows, rather than staying a result that only holds up in a datacenter.