Technical: CV production, DDP and TensorRT

Dropout, BCE versus accuracy, BatchNorm in DDP, TensorRT/ONNX optimization, quantization/calibration, detector resolution and perception-team context.

6 questions28 minAudio recording available

Question languageРусский English Оригинал

Пройти собеседование Весь ход интервью

Аудио и материалы

Аудио собеседования

0:00 / 28:19

Этап 1 из 3NavioDeep Learning Researcher, Perception2025-08-06 - 2025-08-15

Собеседования в Navio: компьютерное зрение

Техническое собеседование в Navio

Весь процесс Этап 2

Выводы и как готовиться

The round was practical and production-leaning rather than math-heavy: how training behavior changes at inference, how metrics can disagree, and how deployment optimizers work.
The automotive context matters: model quality is tied to sensor setup, small objects, latency, hardware constraints and safety-critical perception behavior.
The interviewer also described the next loop: a medium LeetCode-style algorithm section and final conversation with the team lead.

Technical: CV production, DDP and TensorRT

Аудио и материалы

Аудио собеседования

Dropout during training and inference

Validation accuracy and BCE loss both increase

BatchNorm problems in distributed training

TensorRT, ONNX, graph optimizations and quantization

Changing input resolution for a YOLO-style detector

Perception team context: 2D/3D detection, fusion, sensors and monthly planning

Выводы и как готовиться