Job description

Future AR and VR devices occupy ever-smaller, glasses-like form factors and thus require ever-tighter power budgets. At the same time, we require higher-quality passthrough rendering and scene understanding. In particular, depth estimation, which reconstructs scene geometry in real time from sensor signals, is one of the biggest power consumers in AR and VR devices. However, most existing work targets datacenter GPUs, or at best mobile silicon with an order of magnitude more power headroom than what glasses form factors allow. Reality Labs is Meta's innovation engine for next-generation AR/VR and wearable technology and XR Graphics is involved in building the next-generation imaging and reconstruction stack that closes this gap. We are looking for a Research Scientist Intern to work on low-power real-time reconstruction of scene geometry for AR/VR. Today, VR passthrough, hand interaction, and environment understanding run on separate pipelines, each tuned for different accuracy, latency and power target. We are interested in deduplicating the compute (and thus reducing power): - What is the right depth estimation and processing strategy, from the ground up, for a glasses power budget? - Can we train a single lightweight scene representation that can serve passthrough, hand occlusions, and background occlusions simultaneously? These are open questions, and you would help define the approach. This is a research internship. You will own a genuinely open problem, choose the methods, build the evaluation, and present your results, with support for publication where the work warrants it. Plan and execute research to advance the state of the art in low-power, real-time scene reconstruction for AR/VR — spanning depth estimation, occlusion, and geometric scene understanding. Explore the extent to which a shared, learned scene representation can serve multiple downstream consumers at different latency and accuracy points. Define evaluation methodology and run experiments measuring geometric accuracy, temporal stability, rendering distortion, latency, and compute and power cost on real multi-sensor device captures. Collaborate with researchers and engineers across the passthrough, scene understanding, and on-device compute teams to assess feasibility within headset and glasses power envelopes. Document and present findings, and contribute to publications and patents where the results warrant it. Currently has, or is in the process of obtaining, a PhD degree in the domain of computer vision, machine learning, computer graphics, robotics, or a related technical field Must obtain work authorization in the country of employment at the time of hire, and maintain ongoing work authorization during employment Knowledge and hands-on experience in at least one of: depth or stereo estimation, 3D reconstruction (TSDF, implicit or neural fields, Gaussian splatting), video or instance segmentation, or neural scene representations Experience designing and running quantitative experiments, and communicating technical results Intent to return to a degree-program after the completion of the internship Proven track record of achieving significant results as demonstrated by grants, fellowships, patents, as well as first-authored publications at leading workshops or conferences such as CVPR/ECCV/ICCV, SIGGRAPH/SIGGRAPH Asia, 3DV, ICLR, NeurIPS, etc Experience with efficient or on-device inference: quantization, distillation, pruning, or deployment to mobile NPU, DSP or GPU targets Experience with multi-sensor fusion across heterogeneous sensors — RGB, IR/CV cameras, time-of-flight — including camera calibration and cross-sensor registration Experience with temporal filtering and state estimation over scene state Familiarity with AR/VR passthrough, reprojection, occlusion rendering, or SLAM and scene reconstruction Experience with C++, and with profiling or optimizing real-time computer vision or graphics pipelines Experience working in a Unix environment Experience reasoning about compute and power budgets as a first-class design constraint

Requirements

Must-haves

What is the right depth estimation and processing strategy, from the ground up, for a glasses power budget?
Can we train a single lightweight scene representation that can serve passthrough, hand occlusions, and background occlusions simultaneously?

Nice-to-haves

None flagged.

Application prep

Use the role’s language and application details while tailoring your materials.

Keywords to mirror

machine learning

Platform: meta
Effort: medium

Benefits

Benefits detected in the listing.

Vision