Computer Vision Engineer
The role
You build the perception pipelines that turn Trinet egocentric recordings - stereo head cameras, wrist cameras, and IMU - into robot-learning training data: 3D hands in a metric world frame, head trajectories and maps of the scene, and per-hand action labels that capture what really happened, mistakes and corrections included. You'll own problems end to end, from camera geometry and calibration, through models running on the GPU, to a label a reviewer can trust - and you'll measure every change for accuracy, speed, and cost.
What you'll get to work on
- 3D hand tracking from monocular and stereo fisheye video: keypoints, wrist 6-DoF, and meshes that hold up through occlusion, grasps, and hands leaving the frame.
- Visual-inertial SLAM for head-mounted cameras: stereo and mono VIO, rolling-shutter correction, camera-IMU self-calibration, loop closure, stereo depth, and dense maps.
- Putting hands, head, and scene into one metric world frame, synchronised across head and wrist cameras.
- Action segmentation and labelling: kinematic cues (grip, motion, contact) combined with vision-language models, with verification, confidence scores, and a human review queue.
- Making it fast and cheap: GPU throughput, batching, mixed precision, and torch.compile, so hours of video become labels in minutes.
- Benchmarking against public datasets and our own ground truth, and turning new papers into working experiments within days.
What we're looking for
- Solid multi-view geometry: camera models (including fisheye), calibration, stereo, triangulation, PnP, and coordinate-frame transforms.
- A solid grip on the math underneath - linear algebra, 3D rotations, optimization, and filtering / state estimation.
- Hands-on 3D vision experience - pose estimation, SLAM / VIO, or depth - ideally on real, messy data, not only benchmarks.
- Strong Python and PyTorch; comfortable reading, building, and patching C++ libraries.
- Knows how cameras and IMUs really behave: rolling shutter, exposure, motion blur, timestamps, and clock offsets.
- You measure before you claim: you build the evaluation first, and you report the numbers that got worse too.
Bonus points
- Hand or body pose estimation, or hand-object interaction research.
- Egocentric vision, robot learning, VLA models, or robot datasets such as LeRobot.
- VLMs / LLMs in production pipelines: structured outputs, evaluation, fine-tuning and serving open-weight models.
- Segmentation and depth foundation models (SAM, Depth Anything) and learned stereo.
- GPU performance work: profiling, CUDA, TensorRT.
- 3D visualisation tools: Rerun, Open3D, three.js.
- Knowing which models and datasets can be used commercially, and checking before building on them.
How to apply
Email hiring@panoculonlabs.com with the subject Computer Vision Engineer - Your Name and answer one question, in your own words, in under 100 words:
What's something you built that you're proud of, and what went wrong while you were building it?
Include a link or a photo that shows the thing you built. That's it. We'll ask about everything else if we reply.
Write it yourself. Short, plain sentences and typos are fine. Emails that read as AI-written, run past 100 words, or are only a resume are rejected without a reply.
Thank you for taking the time to apply. We read every email, but because of the number of applications we receive, we're only able to reply to people we'd like to speak with. If you don't hear back from us, it means we won't be moving forward this time. We truly appreciate your interest and wish you the best.
Email hiring@panoculonlabs.com