Some more observations. I've been trying to make the EOS-M a nice portable RAW camera (720p), as an alternative to the BMPCC. One of the obstacles are the focus dots that appear in the raw frames. They generally turn into pink dots after debayering (because most are in the red pixels). Someone created "PInk Dot Remover" in Java which is certainly a very slick and effective piece of software However, it doesn't get all the dots (interpolate around the "hot" focus pixels). The first dev is no longer around, nor the second guy who took it up. I also wanted something in Windows, just to make my workflow easier.
Anyway, I use a C# project open-sourced by g3gg0 (one of the lead rock stars) to get me going. It has a function to read pixels from the RAW files, but not to write them. I ask him if he can do it and he says he's too busy. No problem. I figure it can't be too hard. Well, it takes me hours upon hours, and I can't do it.. I finally go back to him and another developer and ask them what the format of the RAW data is. They don't seem to know! Seems incredible to me. How can they not know the basic format of pixel data in RAW files? I believe the answer is that they took routines from elsewhere and modified them a bit without knowing exactly what these routines do (lots of bit-wise arithmetic). They were able to get each pixel's value into a ushort and that was it. As smart as they are, they are overloaded by technical stuff, which I believe even they can't get completely on top of.
I finally figured out that the data was encoded as a 14-bit stream where each other byte was swapped with the one before it. If I had known the original format I could have figured this out easier. The point of this story is that as much as the devs seem to know, the stuff is monstrously difficult on so many levels. Knowing what I know now, I'm amazed ML continues to move forward. However, if one or two of those key guys quite that would be the end, or severely slow down the development of ML. I believe this is not unlikely. Because...
Video is a narrative medium. Yes, it's fun to take some RAW video of your garden, kid, tugboats, etc. After that, no one wants to watch a video that doesn't tell some kind of story. Story requires scripting, editing, actors, lighting, audio, etc., etc. I don't know if this is heresy, but I can't see why someone would use a 5d3 to shoot a film when they could use a BMCC or any of the other dedicated cameras for that purpose. Now that the BMPCC seems to get in stock I"m thinking, I have $850 in a used 50D (CF card+10-20mm) to shoot RAW, what am I thinking! I've already watched 2 people on the ML forum sell their 50ds, or try to. I may be next. The other thing about narrative video, VERY FEW people actually do it. That is a VERY SMALL market. And many of them, like Andy, don't have the time or need for any kind of RAW-type workflow.
Yes, I believe the 5D3 is a better value than the BM cameras, but only if you're going to spend the time/effort to get the most out of the RAW feed. If you just need basic improved dynamic range, the BM cameras make so many things easier. And they will continue to improve.
Because the devs won't try to make ML stable on these cameras, or even allow the general public discussion of a to-do list, I can't see it growing beyond a "cool" hack. Maybe I'm too close to it. Familiarity is starting to breed contempt ;)