Year One at Loopbio

A year ago I wrote, in Another Change , that I was leaving academia to build the tools I had wanted as a researcher. This is what the first year of Loopbio involved.

Most of it was hardware. We built camera-synchronisation boards, tested in arrays of ten because the systems we sell often use that many; frames in black anodised T-slot extrusion; and lighting rigs. Lighting is more work than it sounds: most tracking and image-processing problems are, underneath, lighting problems.

I spent a few days at the VISION fair in Stuttgart, looking at 10GigE and NBASE-T cameras (2.5G and 5G over more than ten metres of cable, which solves a real problem for us), lenses, and 3D and time-of-flight imaging. I also read the NIPS write-ups, and we bought a Titan X Pascal to build up the compute for deep learning. Deep learning is a useful tool, and we are investing in it without treating it as a magic one.

The main work was with the scientists. Simon Gingins sent us large volumes of collective-behaviour footage of fish from the Red Sea, off Eilat, and we built wand autocalibration so that the behaviour could be reconstructed in three dimensions rather than from a single camera. There was fish VR with the collective-behaviour group, zebrafish tracking with real-time object detection, and 3D cup tracking built up calibration sphere by calibration sphere. They get on with the science while we handle the apparatus, which is the point of the company.

It was a year of more customer support and more building than the original plan implied. Thanks to the collaborators who trusted a one-year-old company with their data, and to Max.