Tommy Flanagan, the editor of Faultline, looked at our ML-Safe work and how we are validating it for the autonomous vehicle market - the storage and networking challenge and what it takes to make massive video datasets smaller without changing what the models see.

"Extending beyond its traditional M&E customer base, Beamr is well on the path towards computer vision and, ultimately, autonomous vehicles. Beamr provides ML-Safe video data compression technology designed to reduce the massive storage and networking costs generated by autonomous vehicle (AV) test fleets".

"Beamr’s researchers described how a few altered pixels might be invisible to a person but could potentially change the output of an object detector, lane detector or vision-language model. A compression artefact that nobody notices while watching a road scene could, in theory, influence where an algorithm thinks a vehicle, pedestrian or lane marking is located.".

"Obviously there is no universal compression setting at which every AI model remains unaffected. Different models react differently to changes in the underlying video, so Beamr is trying to establish a methodology for measuring that sensitivity before a compression regime is deployed".

Read the article here