Do Computer Vision companies use OpenCV as their main tool ?

I've worked at two different startups and JPL, so this may not be a fair analysis as in these cases quick iteration and deployment is more important than micro-optimization. For basic stuff like reading images, morphology, color space conversions, etc, you really can't go wrong with OpenCV and trying to improve performance in any of these realms is almost certainly company money wasted. OpenCV's complete algorithm implementations are also typically good enough for classical algorithms, although I've never used and couldn't recommend any of their DNN modules over PyTorch. From here, I'll either edit source or replace modules as necessary - for example, if matching is a bottleneck in computing homographies and exact matches are required, pulling a VP tree implementation could help. I've recently used their Line2D implementation, which benefited from inserting multithreading and SIMD.