Research Scientist: 3D Vision & Egocentric Motion

Kindredventures; New York (NY); On-site; USD 150,000 - 210,000; Full time

Job summary

Mecka AI in New York seeks a Research Scientist to architect and train foundation models from scratch for spatial-temporal perception in embodied AI. You will lead development of 3D geometry extraction networks and egocentric optical flow models using proprietary ground-truth data.

You will scale training across multi-GPU clusters, explore novel losses and CUDA-accelerated ops, and advance dense reconstruction techniques (NeRFs) for photorealistic environments, while adapting quickly to evolving

Qualifications

• Deep knowledge of modern deep learning, 3D vision, and temporal/video modelling.
• Proven track record training large-scale vision models from scratch.
• Strong understanding of feed-forward 3D networks and dense motion estimation.
• Mastery of PyTorch and scalable DL frameworks.
• Experience handling multi-terabyte image/video datasets.

Responsibilities

• Architect and train proprietary spatial models from scratch.
• Scale multi-GPU distributed training for multi-view datasets.
• Develop novel losses and attention mechanisms for 3D data distributions.
• Build egocentric optical flow models and temporal dynamics for dynamic scenes.
• Prototype and deploy models for tracking, segmentation, and fusion in robotics contexts.
• Integrate models into dense reconstruction pipelines (e.g., NeRFs).

Skills

Deep learning

3D vision

Temporal modeling

PyTorch

Large datasets

Adaptability

Tools

CUDA

Ray marching

Correlation volumes

Job description

Mecka AI in New York seeks a Research Scientist to architect and train foundation models from scratch for spatial-temporal perception in embodied AI. You will lead development of 3D geometry extraction networks and egocentric optical flow models using proprietary ground-truth data.

You will scale training across multi-GPU clusters, explore novel losses and CUDA-accelerated ops, and advance dense reconstruction techniques (NeRFs) for photorealistic environments, while adapting quickly to evolving