About the Team & Mission

42dot의 Computer Vision Engineer는 자율주행을 위한 통합 비전 모델과 학습 기술을 연구·개발합니다. 대규모 주행 데이터에서 3D 객체 검출, 도로 구조 인식, occupancy 등 3D scene understanding 과제를 함께 학습하고, 여러 GPU와 노드에 걸친 분산 학습의 효율과 안정성을 높입니다. 또한 실제 주행과 평가에서 발견되는 성능 저하의 원인을 분석하고, representation learning과 self-supervised learning을 통해 다양한 주행 환경에 일반화할 수 있는 시각 표현을 개발합니다.

As a Computer Vision Engineer at 42dot, you will research and develop unified vision models and learning methods for autonomous driving. You will train models across 3D scene understanding tasks, including 3D object detection, road structure understanding, and occupancy prediction, using large-scale driving data, while improving the efficiency and stability of distributed training across multiple GPUs and nodes. You will also investigate performance regressions observed in real-world driving and evaluation and develop visual representations that generalize across diverse driving environments through representation learning and self-supervised learning.

Responsibilities

1. Multi-Task Learning & Distributed Training Optimization

Multi-view geometry와 3D scene representation에 대한 이해를 바탕으로, 3D 객체 검출·도로 구조 인식·occupancy 등 다양한 자율주행 인식 과제를 통합하는 multi-task learning 모델과 학습 전략 연구·개발

Task 간 상호작용을 분석하고, 데이터 구성·sampling·loss balancing·학습 스케줄을 개선하여 개별 task 성능과 통합 모델 성능 향상

대규모 multi-GPU·multi-node 학습의 연산, 메모리, 통신 및 데이터 파이프라인 병목 분석·최적화

Research and develop multi-task learning models and training strategies that integrate autonomous driving perception tasks, including 3D object detection, road structure understanding, and occupancy prediction, grounded in multi-view geometry and 3D scene representation.

Analyze interactions between tasks and improve data composition, sampling, loss balancing, and training schedules to enhance both individual task performance and overall model performance.

Identify and optimize computation, memory, communication, and data pipeline bottlenecks in large-scale multi-GPU and multi-node training.

2. Regression Analysis & Model Improvement

모델 변경과 실제 주행 평가에서 발견되는 성능 저하 및 실패 사례를 재현하고, 정량·정성 분석을 통해 원인 규명

데이터·GT 품질, 전처리, 모델 구조, 학습 조건 및 배포 과정의 영향을 구분하는 가설 수립과 ablation 실험 수행

Reproduce performance regressions and failure cases identified through model changes and real-world driving evaluation, and determine their causes through quantitative and qualitative analysis.

Formulate hypotheses and conduct ablation studies to distinguish the effects of data and ground-truth quality, preprocessing, model architecture, training conditions, and deployment.

3. Pre-training & Representation Learning

대규모 비라벨 주행 데이터를 활용한 representation learning 및 self-supervised learning 기법 연구·개발

Self-supervision, knowledge distillation, auxiliary supervision을 결합한 multi-view·temporal 정보와 기하학적 관계를 활용하는 pre-training recipe 설계·검증

Representation collapse와 학습 불안정성을 분석하고, 학습된 표현의 품질 및 downstream task로의 전이 효과 평가

Research and develop representation learning and self-supervised learning methods using large-scale unlabeled driving data.

Design and validate pre-training recipes combining self-supervision, knowledge distillation, and auxiliary supervision.

Investigate representation collapse and training instability, and evaluate representation quality and transfer to downstream tasks.

Qualifications

컴퓨터비전 및 딥러닝에 대한 깊은 이해와 2년 이상의 모델 연구·개발 경험

3D geometry, 카메라 모델 및 좌표 변환, multi-view geometry에 대한 이해와 이를 활용한 3D vision 모델 개발 경험

Multi-GPU 분산 학습 경험과 학습 성능·메모리·수치적 안정성 문제를 분석하고 개선한 경험

Multi-task learning, representation learning, self-supervised learning 중 하나 이상에 대한 심도 있는 연구 또는 개발 경험

A deep understanding of computer vision and deep learning, with at least 2 years of experience researching and developing models.

An understanding of 3D geometry, camera models, coordinate transformations, and multi-view geometry, with experience applying these concepts to 3D vision model development.

Experience with multi-GPU distributed training and diagnosing and improving training performance, memory usage, and numerical stability.

In-depth research or development experience in at least one of multi-task learning, representation learning, or self-supervised learning.