Computer Vision Engineer (Autonomous Driving)

About the role

42dot is hiring a Computer Vision Engineer in Pangyo to research and develop unified vision models and learning methods for autonomous driving. You will work on 3D scene understanding tasks such as 3D object detection, road structure understanding and occupancy prediction, using large-scale driving data, and improve the efficiency and stability of distributed training across many GPUs and nodes. You will also analyse performance regressions seen in real driving and build visual representations that generalise across environments.

Key responsibilities

• Research and develop multi-task learning models and training strategies that combine 3D object detection, road structure understanding and occupancy prediction, and optimise data composition, sampling, loss balancing and training schedules.
• Analyse and optimise bottlenecks in computation, memory, communication and data pipelines in large-scale multi-GPU and multi-node training.
• Reproduce performance regressions and failure cases found in real driving, and run ablation studies to isolate the effect of data quality, preprocessing, model architecture and training conditions.
• Develop representation learning and self-supervised learning methods on large-scale unlabeled driving data, design and validate pre-training recipes, and evaluate representation quality and transfer to downstream tasks.

Required profile

• Deep understanding of computer vision and deep learning, with at least 2 years of experience researching and developing models.
• Experience with multi-GPU distributed training and fixing issues of performance, memory and numerical stability.
• In-depth research or development in at least one of multi-task learning, representation learning or self-supervised learning.

Preferred qualifications

• Publications at top-tier conferences in computer vision, robotics or machine learning.
• Experience improving scaling efficiency and GPU utilisation in large-scale multi-node training.
• Experience solving multi-task learning challenges such as loss balancing, task interference and negative transfer.
• Experience building BEV- or PV-based 3D scene representations from multi-camera and temporal data.

The process includes an application screening, a first virtual interview of about an hour, a second interview of about three hours, and then an offer discussion and onboarding. A 3-month probation period may apply.