About the role
Turn images and video into carefully evaluated research experiments. Work on visual understanding, annotation quality and the evidence behind a model’s results.
What you’ll work on
- Prepare image and video datasets, including frame sampling, annotation checks and reproducible preprocessing.
- Build baselines for visual tasks such as detection, segmentation or temporal understanding, according to the research question.
- Evaluate predictions on held-out examples and inspect failures across lighting, motion and occlusion.
- Keep related frames and videos in appropriate dataset splits to avoid misleading evaluation results.
- Document experiments, data provenance and limitations so another person can reproduce the work.
Core skills
- Python, NumPy and OpenCV for image processing and video handling.
- Working knowledge of PyTorch or TensorFlow and the fundamentals of neural networks.
- An understanding of image coordinates, resizing, colour spaces and frame rates.
- Familiarity with detection or segmentation and metrics such as precision, recall, IoU or mAP.
- Careful annotation, dataset splitting, error analysis and experiment documentation using Git.
Useful additional experience
Helpful, but not required.
- Object tracking, temporal models or video transformers.
- Annotation tools and dataset formats such as COCO.
- Reproducible research notebooks and inference profiling.
Where this work connects
Atlas is research in progress on surgical-video review and assessment. This track would support visual and temporal research experiments; it does not imply a clinically validated product or deployment.
Show us your work
Share an image or video project with a repository or notebook. Describe the dataset, your contribution, evaluation split, chosen metrics and a failure case. Public or permissioned data is sufficient; do not share confidential recordings.
Tell us what you contributed, what you learned and what you would improve. Personal projects and coursework are welcome.
A question to consider
A model scores well on held-out frames but fails on a new video. What would you investigate in the data split and evaluation?
About Ponens Labs
We develop our own products and provide software engineering and cybersecurity services. Our work connects product thinking, implementation and careful review.
Learn about Ponens