[M4IF Technotes] Coder efficiency

[Yang, Wei Jian (Mac)]

Dear MPEG4 expert:
I think I have find the reason why Momusys generate less bits in P and B
frames.
There is a SAD calculation in mot_est_mb.c:
If calculate sad with MB in same position of previous picture
sad=SAD_Macroblock(...) - (MB_Nb[currentMB]/2+1);
else
sad=SAD_Macroblock(...);
MB_Nb[...] is number of pixels of MB in shape. If I use rectangle coding,
MB_Nb[...] should be 255.
So, MB at MV(0,0) has much more previlige than other candidate MBs.
My question is, sad should be like the following to make more sense:
sad=SAD_Macroblock(...) -(255-MB_Nb[currentMB]);
Who is correct?
Thanks
Mac
-----Original Message-----
From: Yang, Wei Jian (Mac) [mailto:yangwj   lucent.com]
Sent: 2002Äê10ÔÂ10ÈÕ 14:38
To: Technotes (E-mail)
Subject: [M4IF Technotes] Coder efficiency
Dear MPEG4 experts:
When I use Momusys as reference MPEG4 codec, I find that something
interesting.
This DCT-like coder has much better efficiency than MPEG2
Simulating environment is:
1. MPEG2 coder: 1996, MPEG Software Simulation Group, MP   ML
2. MPEG4 coder: MoMuSys-FDIS-V1.0-990812 simple profile
3. Akiyo.yuv
4. 96000kbps, MPEG2 TM5 Rate Control
Compare result is: (bits/PSNR-Y(db))
	    I		P	B	 ......
MPEG2	37740/32         7234/32.6   9979/33.1
MPEG4	47982/37.8      678/37.5     80/37.7
I find that MPEG2 coder produce much more MVs in P and B than those in
MPEG4.
I guess in MPEG4, it uses a more efficient MB type decision policy in Motion
Estimation.
I know MPEG4 adopts quarter-pixel prediction, but does it has such high
impact on
ME efficiency? 
Any comments?
Thanks
Mac
Yang, Wei Jian (Mac)
Bell labs, Shanghai
Lucent Technologies (China) Co., Ltd.
Tel:         8621-5450-4555  ext. 8610
Fax:        8621-5490-2106
E-mail:   yangwj   lucent.com
_______________________________________________
Technotes mailing list
Technotes   lists.m4if.org
http://lists.m4if.org/mailman/listinfo/technotes