How to improve the quality of the h264 encoder?

[Lior_Messinger]

Hi

We are looking for a way to improve the quality of the h264 encoder, which is suboptimal. except the deblocking filter, which produces a good smoothing, we cannot find anything else which provides a real improvement. could anyone suggest anything?

here's a visual example. (B) supposed to be much superior to (A), but still, we cannot really see ANY difference.

any ideas? any explanations?

thanks much

Lior

A. http://img33.imageshack.us/img33/5813/low400kbpsdeblockingh26.jpg

this image uses baseline profile (66), deblocking filter and

2 1 1 /* num_ref_frames (2-16), minimum length of list1 for backward prediction , number of slices. */

3 0 4 4 /* ME method (1-6), subblock split, search x,search_y */

1 0 /* direct type (0 - temporal 1 - spatial 2 - auto); direct_inference_flag */

2 /* picture coding type */

B> http://img638.imageshack.us/img638/8797/low400kbpslatesthh264hi.jpg

this one is when using high profile, deblocking and

8 1 1 /* num_ref_frames (2-16), minimum length of list1 for backward prediction (only 1 is supported!), number of slices. */

2 2 12 12 /* ME method (1-6), subblock split, search x,search_y */

1 1 /* direct type (0 - temporal 1 - spatial 2 - auto); direct_inference_flag */

0 /* picture coding type */

Hi
We really need a way to optimize the quality of the resulting video, which is suboptimal. except the deblocking filter, which produces a good smoothing, we cannot find anything else which provides a real improvement. could you suggest anything?
here's a visual example. (B) supposed to be much superior to (A), but still, we cannot really see ANY difference.
any ideas? any explanations?
thanks muchLior
A. http://img33.imageshack.us/img33/5813/low400kbpsdeblockingh26.jpgthis image uses baseline profile (66), deblocking filter and2 1 1 /* num_ref_frames (2-16), minimum length of list1 for backward prediction , number of slices. */3 0 4 4 /* ME method (1-6), subblock split, search x,search_y */1 0 /* direct type (0 - temporal 1 - spatial 2 - auto); direct_inference_flag */2 /* picture coding type */

B> http://img638.imageshack.us/img638/8797/low400kbpslatesthh264hi.jpgthis one is when using high profile, deblocking and
8 1 1 /* num_ref_frames (2-16), minimum length of list1 for backward prediction (only 1 is supported!), number of slices. */2 2 12 12 /* ME method (1-6), subblock split, search x,search_y */1 1 /* direct type (0 - temporal 1 - spatial 2 - auto); direct_inference_flag */0 /* picture coding type */

[dzhou_tac]

Hi,

I found there are problems with IPP H264 Encoder, which creates much bigger encoded size than original h264 stream.

Have you try to increase bitrate? This may generate better image. I know to increaes the bitrate probably is not you are asking.

david

[andrewk88]

Quoting Lior Messinger

Hi

We are looking for a way to improve the quality of the h264 encoder, which is suboptimal. except the deblocking filter, which produces a good smoothing, we cannot find anything else which provides a real improvement. could anyone suggest anything?

here's a visual example. (B) supposed to be much superior to (A), but still, we cannot really see ANY difference.

any ideas? any explanations?

thanks much

Lior

A. http://img33.imageshack.us/img33/5813/low400kbpsdeblockingh26.jpg

this image uses baseline profile (66), deblocking filter and

2 1 1 /* num_ref_frames (2-16), minimum length of list1 for backward prediction , number of slices. */

3 0 4 4 /* ME method (1-6), subblock split, search x,search_y */

1 0 /* direct type (0 - temporal 1 - spatial 2 - auto); direct_inference_flag */

2 /* picture coding type */

B> http://img638.imageshack.us/img638/8797/low400kbpslatesthh264hi.jpg

this one is when using high profile, deblocking and

8 1 1 /* num_ref_frames (2-16), minimum length of list1 for backward prediction (only 1 is supported!), number of slices. */

2 2 12 12 /* ME method (1-6), subblock split, search x,search_y */

1 1 /* direct type (0 - temporal 1 - spatial 2 - auto); direct_inference_flag */

0 /* picture coding type */

Hi
We really need a way to optimize the quality of the resulting video, which is suboptimal. except the deblocking filter, which produces a good smoothing, we cannot find anything else which provides a real improvement. could you suggest anything?
here's a visual example. (B) supposed to be much superior to (A), but still, we cannot really see ANY difference.
any ideas? any explanations?
thanks muchLior
A. http://img33.imageshack.us/img33/5813/low400kbpsdeblockingh26.jpgthis image uses baseline profile (66), deblocking filter and2 1 1 /* num_ref_frames (2-16), minimum length of list1 for backward prediction , number of slices. */3 0 4 4 /* ME method (1-6), subblock split, search x,search_y */1 0 /* direct type (0 - temporal 1 - spatial 2 - auto); direct_inference_flag */2 /* picture coding type */

B> http://img638.imageshack.us/img638/8797/low400kbpslatesthh264hi.jpgthis one is when using high profile, deblocking and
8 1 1 /* num_ref_frames (2-16), minimum length of list1 for backward prediction (only 1 is supported!), number of slices. */2 2 12 12 /* ME method (1-6), subblock split, search x,search_y */1 1 /* direct type (0 - temporal 1 - spatial 2 - auto); direct_inference_flag */0 /* picture coding type */

If I'm not mistaken you're trying to encode a relatively high resolution video at low bitrate "400kbit/sec", so in general your expectations for a video quality should not be too high.

Some ideas to consider:

a) try pre-processing of input video

b) set-up more agressive deblocking filtering using offsets

c)try to use these tools:8x8 transform, CABAC, B-frames, multiple ref. frames, don't use INTER sub-partitions smaller than 8x8, use hpel and qpel accurate motion vectors

d) use good rate control scheme

Best regards,

AndrewK

[Lior_Messinger]

Andrew,

Thank you SO MUCH for this reply. this is a huge help, and we are going to try all the above suggestions. We already improved much our pre-processing (resizing, deinterlacing) in recent days, and I wondered if you could elaborate a little on the other suggestions.

specifically, when we looked on h264.par we couldn't find an exact wording that would point to deblocking offsets and half/quarter pixel motion estimation. could you point us maybe?

also, what would be a good rate control scheme? we see in h264.parRC method(0 - quant_codes, 1 - CBR MBwise, 2 - CBR framewise, 3 - Debug) - what do you think we should use? what would be good "start qp values for I, P, B slices" that would improve quality?

I guess we sound like total newbies - which we are, using the h264.par as-is. we are learning on the go, and it just take time to read the books :) (but we are).

in any case, again, very grateful for your help,

Lior

[Chao_Y_Intel]

Hi Lior,

For Rate control method, you can choose 2, here are some sample par files for H.264 encoding, it may provide some help:

http://software.intel.com/en-us/articles/setting-h264-encoding-parameters-in-intel-ipp-media-processing-samples/

Thanks,

Chao