[bitL]

It's pretty common to ramp up contrast in cam used for self-driving inputs for basic computer vision to make lane markers pop up. So such a camera won't have much chance of detecting a person casually crossing road without any concerns for traffic. The crash points more to failure in sensor fusion; LiDAR should have detected it, but clustering algorithm might have either removed it, or she appeared in too few scanning frames to be reliably detected as an object, due to heat map thresholding to avoid false positives.

[ksk]

But why would they alter the recorded bits? An imaging sensor streams the RAW data (w/o tone curve). Wouldn't they prefer just storing the RAW bits (which probably have a higher dynamic range anyway) and then apply various transforms - contrast, edge-detect, etc.

[bitL]

I guess it's speed concerns? Maybe they have a custom ASIC they use to pipe and adjust images from the camera, achieving optimal contrast? Using CPU for that is very very slow and self-driving car companies try to push latency down as much as possible, so it might be a limitation of specialized hardware.

[ksk]

Sure, but if they have to process the RAW data anyway, they could easily send an unaltered copy in parallel to storage for offline use. I guess depending on the resolution & bitness, they'd want to downsample/compress it to save space. Heh, anyway its just pure speculation on my part. They probably have a good reason to do it one particular way...