X264 - Motion Estimation Method- Comparison

Last time we looked at the effects of the various subpixel estimation complexity, today we will be looking at how the motion estimation method(MeM) impacts the filesize(all measured in KB), quality and performance(average FPS) according to CRF18 of x264.

There are a total of 5 methods these are in order of complexity:

• Diamond(dia) - The most simplest method, checking motion vectors at one pixel up, left, down and right and picking the best candidate, this process repeats until it can't find a better motion vector
• Hexagon(Hex) - Work similarly like Diamond but consists of a range-2 and 6 surrounding points, more efficient than diamond with little performance impact.
• Uneven multi-hexagon(Uhm) - While slower than Hex, it is able to avoid missing harder-to-find motion vectors. The Me-Range parameter controls its search radius.
• Exhaustive(Esa) - An optimised intelligent search of the complete motion vector space within Me-Range. Mathematically equivalent to brute force but faster, but still slower than Uhm.
• Transformed exhaustive(Tesa) - is an algorithm which attempts to approximate the effect of running a Hadamard transform comparison at each motion vector; like exhaustive.

So now that we are more familiar with MeM lets see how it performs.


Futurama SD

The top graph shows the FileSize(Graph area) in KB as well as the performance(line) of each of the different methods. As you can see the relationship between compression and performance takes a turn for the worst with Esa and Tesa.

The bottom graph represents the difference in percentage for speed and compression compared to the default(medium) setting. Here we can see that with Esa for 29.5% performance reduction we get a measly 2.56% extra compression and Tesa does even worse. Uhm compresses better with a slight performance hit of 6.71% slower. Of course will have to see if this translates into any visual artefacts

Still Frame

Diamond

Hexagon

Uhm

Esa

Tesa

Even though I can see a VERY slight difference between the different methods, I cannot say that one looks better than the other. 


Smaller Motion Detail

Diamond

Hexagon

Uhm

Esa

Tesa

Again I can notice a difference between the methods but cannot say which is actually higher quality than the rest.

Reel Steel SD
 

So moving onto live action video, we can see that the same trend repeats it self with esa and tesa performing very poorly when compared to dia, hex and uhm.


Still Frame

Diamond

Hexagon

Uhm

Esa

Tesa

As with Futurama its difficult to tell which is better quality.


Smaller Motion Detail

Diamond

Hexagon

Uhm

Esa

Tesa

  .

 

 

And yet again the frame quality is too consistent to say which is better. Maybe we'll see a bigger deference with HD videos

 

Reel Steel HD 

So looking at HD video we see that esa and tesa performs worse than Uhm and is significantly slower. Hexagon seems like a good default method.


Still Frame 

Diamond

Hexagon

Uhm

Esa

Tesa

 

 

 

 

 

 

 

 

 

Yet again with MeM and with HD I can hardly see any difference let alone which is higher quality   

Smaller Motion Detail

Diamond

Hexagon

Uhm

Esa

Tesa

 

So yet a again its almost impossible to tell which is which and which is better than the other.

Conclusion

Well as you can see visually there isn`t much of a difference, in terms of compression Uhm did consistently better than hexagon or diamond, and esa, tesa also performed better except at HD videos. Regarding performance Esa and Tesa took a massive performance hit, with Uhm only being slightly slower.

Recommended
Hexagon - Performed consistently close to the others.
Uhm - A slight compression gain at a margin performance loss.


 Navigation

1. Introduction
2. Presets
3. Subme  
4. Motion estimation method

X264 - Subme - Comparison

x264 Settings Comparison Part : 2 - Sub Pixel Motion Estimation

Subme define the subpixel estimation complexity. Higher values are better the question is by how much and what is the trade off.

Levels 1-5 control the subpixel refinement strength.
Level 6 enables RDO(Rate–distortion optimization) for mode decision, with level 8 enabling RDO for motion vectors and intra prediction modes. RDO levels are significantly slower than the rest.

Values available:
0 - Full pixel Only.
1 - Quater pixel(QPel) SAD 1 iteration
2 - Quater pixel SAD 2 iteration
3 - Half pixel(Hpel) on MB then QPel
4 - QPel Always
5 - Multi QPel + bi - directional motion estimation
6 - Rate Distortion(RD) on I/P
7 - RD on all frames(Default, Medium)
8 - RDO on I/P frames
9 - RDO on all frames
10 - QP-RD(Requires trellis=2, aq-mode(Adaptive Quantization Mode )>0 )*
11 - Full RD(No early termination)*

*Update found out that Subme10,11 requires other settings, as noted in post previously, will update post with new data.

Futurama SD

 Ok so lets take a look at the performance of sub pixel motion estimation

As you can see from the above graphs subme:0-3 shows significant performance improvement, with 0 and 1 have a noticeable adverse affect on compression, whilst 2-3 having better compression. However and this is important subme lower than 4 does have visible artefacts and I would not recommend any setting lower than 5.

We can also see no difference in subme 9,10,11 these values might require other additional settings for them to take effect. Safe settings are 5-8.

So lets take a look at the screenshots where you can decide for yourself.
Still Frame   

Subme=0

Subme=1

Subme=2

Subme=3

Subme=4

Subme=5

Subme=6

Subme=7(Medium)

 

Subme=8

Subme=9

Most of the still frames have similar quality, looking closely higher subme retains more detail.

Smaller Motion Detail

Subme=7(Medium)

Subme=0

Subme=1

Subme=2

Subme=3

Subme=4

Subme=5

Subme=6

Subme=8

Subme=9

 

 

 

It is difficult to see the difference between subme settings and animation type videos, however the higher settings do introduce less artfefacts and the image quality is consistently cleaner. 

Real Steel SD

 

As you can see the trend is similar to that of the previous one, with Subme 0-3 consitently faster at a loss of compression and 9,10,11 being identical.

 

Still Frame

subme=0

subme=1

subme=2

subme=3

subme=4

subme=5

subme=6

subme=8

subme=9

 

 

 

subme=7(Medium)

 

 

 

 

 

 

 

 

Looking at non animation type videos we can clearly see a visible difference between the subme settings with subme:0,1,2 retaining significantly less finer detail than the higher settings. From subme=7 is becomes a lot harder to notice the loss of fine detail.