Computer Vision Engineer (Visual Place Recognition)

Farsight Vision

Що робити

Own Visual Place Recognition end-to-end: problem framing, dataset curation, training, evaluation, and deployment into the FSV Platform.
Design, train, and iterate on VPR models and retrieval pipelines — global descriptors, local feature matching, and hybrid approaches as the task requires.
Work with and extend modern VPR / visual localization methods (for example NetVLAD, MixVPR, CosPlace, EigenPlaces, AnyLoc, DINOv2-based descriptors, SuperPoint / SuperGlue / LightGlue-style matching) and know when classical or learning-based methods fit better.
Build robust training pipelines: data prep, augmentations for domain shift, hard-negative mining, loss design, hyperparameter search, and reproducible experiment tracking.
Evaluate models beyond top-line recall — failure modes under night, blur, compression, seasonal change, and map—query domain gap; define metrics that matter for field use.

3+ years in computer vision / deep learning for real products or research transferred to production.
Strong Python and hands-on experience with PyTorch (or equivalent) for training and debugging vision models.
Proven experience with Visual Place Recognition, visual localization, image retrieval, or closely related geo-localization / SLAM-adjacent perception.
Ability to train and fine-tune VPR / retrieval models from scratch or from public checkpoints: datasets, losses, mining strategies, and evaluation protocols (e.g. recall@N, precision—recall under realistic splits).
Working knowledge of modern VPR and matching literature and the judgment to pick methods for viewpoint, appearance, and domain shift — not only recreate paper numbers.

Знайдено через trackr.help/jobs · Канал: @trackrhelp · Бот для персональних сповіщень: @trackrhelpBot