Implement codec - help required.

[DrPhill]

Hi all,

I have started a new approach to my ongoing project anf would appreciate pointers to information to help me.

What I am hoping to do is decode mpg, do some video editing, and recode the mpg. I am working in Java (where I am most comfortable). Now, before you say 'this is a ver big task' let me explain.....

I am only intending to decode one particular format - the one produced by my camera. It actually seems very simple so far: One picture per group, one slice per picture, only I frames, no special intraQuantizeMatrix or non-intraQuantise matrix.

I have gleaned all that from the video stream itself - I have the source code for a java mpg player to guide me. It only copes with level one mpg and my video does not work with it. I think it may be leveltwo with paaccketisation, but I think I have identify the 'packet markers'(??) in the stream, though they seem irrelevant to my purposes. I can identify the various start codes and extract information.

I have now drilled down as far as the slice, and need to start reading macroblocks. I am not sure what the code is doing at this point with the macroblock addresses.

So I need a clear explanation of the MPG format at this point - anyone know where I could find one?

I also believe that with I_frames I will not need DCT stuff - that this is for frame types other than I_frame. Am I correct?

I think I can figure out the macroblock --> image stuff, and then the image to RGB32 so that I can do my editing.

Then, on the far side, I will need to know the bare minimum MPG stream content I need for a well formed mpg. Then I can feed my results to a more competent program for serious encoding.

Anybody have pointers to information I will understand and find useful (if those are not mutually exclusive terms here).

Sorry for the long post, and sorry if this is the wrong forum.

All help gratefully received.

[Guest]

What I am hoping to do is decode mpg, do some video editing, and recode the mpg. I am working in Java (where I am most comfortable). Now, before you say 'this is a ver big task' let me explain..... Your explanation doesn't make it any less of a job.

I have gleaned all that from the video stream itself - I have the source code for a java mpg player to guide me. It only copes with level one mpg and my video does not work with it. I think it may be leveltwo with paaccketisation, but I think I have identify the 'packet markers'(??) in the stream, though they seem irrelevant to my purposes. I can identify the various start codes and extract information. So you need MPEG2 in a transport stream. Why don't you try to port some existing code to Java (if you must, it's a huge undertaking), such as DGMPGDec?

I have now drilled down as far as the slice, and need to start reading macroblocks. I am not sure what the code is doing at this point with the macroblock addresses.

So I need a clear explanation of the MPG format at this point - anyone know where I could find one? Refer to the specs: ISO 13818-1 for transport streams, and ISO 13818-2 for MPEG2 video.

I also believe that with I_frames I will not need DCT stuff - that this is for frame types other than I_frame. Am I correct? Wrong.

I think I can figure out the macroblock --> image stuff, and then the image to RGB32 so that I can do my editing. You know all about IDCT, motion compensation, and all that?

Then, on the far side, I will need to know the bare minimum MPG stream content I need for a well formed mpg. Then I can feed my results to a more competent program for serious encoding. You're going to make an MPEG2 encoder too?

All help gratefully received. Don't do it! Try to make a workflow with existing tools.

[DrPhill]

Always with the negative waves, Moriarty..... :D

I am sure that you meant your advice kindly, bI learn by doing, neuron2. I have learnt quite a bit so far about (this one) video stream format, though the task does seem a little challenging.

- DCT are no more frightening than FFT (to which they are closely related) and of which I have coded several. I had not realised they were used for compression, thinking rather that they were used for predictive frames. A bit more learnt.

Since I only have I frames, no motion compensation is needed.

I have converted existing decoder to use level two (of the transport rather than mpg?) and it copes well with my videos I now have my source of images to manipulate. The difference between levels one and two turned out to be trivial - the packetisation of the video data (to intersperse it with audio?).

I do not imagine that I will get so lucky as to find a coder written in Java when I need to export my data, but I may be able good stab at a readable format since I now have a reader to test with. rDCT will be the trickiest bit. I probably do not need to use vlc as long as I avoid reserved values - I will let another program figure out maximum compression. Similarly, motion compensation will not be needed as I can write I frames. Still a large task, I admit, but not as large as some I achieved in a professional capacity (solving this kind of challenge is fun - and it made me enough to retire in 15 years professional work).

If the worst comes to the worst then I can output a collection of stills for another program to stitch together, but for now I can do the fun bit....... video manipulation.

[kalehrl]

Have a look at this Java-based MPEG2 demuxing/editing tool:
http://sourceforge.net/projects/project-x/

[DrPhill]

An interesting looking project. Thank you.

[DrPhill]

OK, so some of this is easy enough. I have implemented and proved the following layers:

RGB32 <==> Y'CrCb
Y'CrCb <==> MacroBlocks
MacroBlocks <==> DCT Blocks
DCT Blocks <==> Quantised Blocks
Quantised Blocks <==> ZigZag blocks

But I am having a little difficulty with the Huffman/LZ77 step. I understand the theory, but do not know which values/codes are used in the standard encoding/decoding table. Do I really need to buy the standards document?

Is there an open source implementation of this portion of the code somewhere that I can read to understand? I can cope with almost any implementation language, but prefer C, C++, or best of all (of course), Java. A source for an implementation may also be optimised in ways that I would struggle to discover.

Working up from the botom of the stack, I have a bitWriting class and can write the MPG stream validly - apart from the picture blocks. I prove this by stealing the picture block code from a real MPG and inserting them to get something playable. eg: I can now code a program to reverse a video and write it to disk in a format that a standard player can render.

Any help or pointers that you can give me would be very welcome. I feel that I am only a couple of days work away from completing the encoder. I could try understanding and reversing the logic from the decoder but that may be a much longer task (weeks, instead of days). And it would not yield an efficient encoding algorithm.

[Remember I am only doing primitive mpg with I-Frames. No fancy motion-compensation or predictive stuff]

I have learnt a lot so far........

[DrPhill]

OK, thanks for all the help.

Job done, I think this is about as good as I can expect: http://youtu.be/g6n_gmibG10.
It was harder than I expected, but not too hard; the main problem was building the tools to analyse the stream so as to identify the bugs correctly.

I now have encode/decode MPG1 in Java.

I think I will tackle MPG4 this winter when I have some free time......