We are looking for a Research Scientist to advance the 3D reconstruction technology behind Athlence’s capture and rendering pipeline.
The focus of this role is on reconstructing fast-moving dynamic scenes from sparse multi-view inputs and rendering them from novel viewpoints while maintaining high visual fidelity, temporal consistency, and computational efficiency. You will work with 3D/4D Gaussian Splatting, feed-forward 3D models, and other reconstruction approaches, adapting existing methods and developing new ones for robust, photorealistic reconstruction.
The role sits at the intersection of 3D vision and high-performance systems. You will work closely with researchers in generative modeling and with hardware/systems engineers to connect reconstruction methods with real-world capture infrastructure and develop models that operate reliably on multi-camera sports data.
What you will be doing
- Research and develop 3D/4D Gaussian Splatting and modern feed-forward 3D models for high-quality dynamic scene reconstruction and novel-view synthesis
- Develop methods for reconstructing dynamic scenes from sparse multi-view observations while preserving geometry, appearance, and temporal consistency
- Explore model architectures, scene representations, training objectives, and other approaches to improve reconstruction quality, robustness, latency, and compute efficiency
- Design and run rigorous experiments to understand model behavior, failure modes, and trade-offs between visual fidelity, geometric consistency, runtime, and compute
- Develop evaluation metrics and benchmarks that capture reconstruction quality across geometry, appearance, and novel-view rendering
- Build reliable training, evaluation, and inference pipelines for 3D reconstruction models.
- Depending on your interests, contribute to the data pipeline, including processing, filtering, and curating large-scale multi-camera datasets for training and evaluation.
- Collaborate closely with generative modeling and hardware/systems engineers to connect reconstruction methods with capture infrastructure and the broader rendering pipeline
What you must have
- Strong hands-on experience in deep learning and computer vision, demonstrated through research, industry experience, or a combination of both
- Strong foundations in multi-view geometry, 3D reconstruction, and camera calibration, including the practical challenges of real-world capture systems
- Deep expertise in at least one of the following areas: 3D/4D Gaussian Splatting, feed-forward 3D reconstruction, neural rendering, or novel-view synthesis
- Strong Python and PyTorch skills, with experience building reliable training and evaluation pipelines
- Experience designing experiments, analyzing model behavior, and investigating failure modes
- Ability to take open-ended research ideas from initial hypotheses through implementation, experimentation, and evaluation.
- Ability to work with a high degree of autonomy in a fast-paced, highly technical, and collaborative environment.
- Strong technical communication skills, with the ability to explain findings, assumptions, and tradeoffs clearly.
Nice to have
- Experience with dynamic scene reconstruction, temporal 3D modeling, or 4D scene representations
- Experience training deep learning models across multi-GPU or multi-node environments, including distributed PyTorch training
- Experience optimizing 3D or deep learning models for training and inference, including memory, throughput, and runtime performance
- Publications at leading computer vision, graphics, or machine learning venues such as CVPR, ICCV, ECCV, SIGGRAPH, NeurIPS, or ICML, particularly in 3D vision or closely related areas
- Experience shipping research-driven computer vision systems into real-world or production environments
- Experience with modern generative image or video models such as diffusion or flow-based models
What we offer
- Access to unique, proprietary multi-camera sports data captured in real-world environments
- The opportunity to work on challenging 3D vision problems where research ideas are tested against real production constraints
- High ownership across the full research cycle, from research exploration and experimentation to integration and deployment
- A small, highly technical team where you can directly shape both the research direction and the systems we build
#J-18808-Ljbffr
