Axis Q1604 & Avigilon H3 1MP Bandwidth Tests

No more mid level

Try night shots

Same area as before but tilt them up/down, left/right to see if omitting certain areas make a big deal (i.e., what happens when we take a shot with 40% sky and then shift down to have no sky?)

Also, just thought about maybe taking the eye chart and using that as a "target" for my wide and narrow shots?

you can try it out,

also try out diferent patterns or pictures indoors, i recommend you try a bunch of things indoors first as it's easier to do so in the office than on the road

Good evening,

Good job, my guess is that the numbers are relatively close because the big driver is the digital noise the Q1604 is generating.

Day vs Night. Toggle them to see if it makes a material difference.

You can lower the gain and that will significantly decrease the bandwidth as well as the details captured. However, we've already done a test on that and it's not very realistic as the tradeoff is bad.

I have the compiled the outdoor bandwidth results here, and I uploaded all of the shots onto dropbox. I want to talk to Ethan about what it is I want to look at for the I-Frame intervals, and whatever else looks important to note with the AVInaptic results.

** For the night shot, I took Lux measurements, and can post those if you would like to see them.

what is the I frame interval for each camera? avinpatic should show it, no?

ok, you updated with the i frame interval, just round to the nearest whole number, i.e., 1.03 is unnecessary, just say 1, also i frame is fixed in almost all manufacturer's implementation so unless you change it explicitly, it is a safe assumption that it is a constant

More importantly, I created a new spreadsheet to match up the numbers for the 2 cameras head to head. It seems one is a night shot and that really screws things up. I created two tables - one with and one without (i.e. just the daytime) because that night time shot has the confounding factor of Axis's super gain / digital noise

Here's the money shot:

Notice that for the same shots, the Axis narrow view is ~2x higher than wide but Avigilon is essentially equal. This despite i frame and DRF being equal.

Either we have screwed something up or we have potentially uncovered something major about differences in encoding. Ethan take a look at this with fresh eyes. We need to do some further investigation here.

Another interesting pattern - less sky did not spike as much as I would have thought. Not sure what is happening there.

Btw, we might want to reach out to Bengt at Ambarella to get a theory from him. Not yet but it could be a next step after we review this more internally.

Question: Do some colors require more bandwidth? Each pixel stores a color value, at least in a keyframe, right?

What I am asking: In the (simplistic, non-hex) example above, does an olive color '0000' consume less bandwidth than cyan '1111'? Doesn't compression work to simplify null bits where possible?

After looking at the Avigilon camera in the office this morning, I think sharpness has something to do with it. We pointed the H3 1MP at the carpet, and switched from sharpness 0 to sharpness 100. Bandwidth increased by 25%. You could quite clearly see more of the pattern on the carpet with sharpness increased.

By looking at the photos, it looks like the Q1604 is much sharper, first off, so patterns on asphalt, leaves on trees, blades of grass, even clouds all are more pronounced. Couple that with WDR (which the Avigilon does not have), and you're reveal (even marginally) more areas. So all of these things being more pronounced equals more "edges" in the image, which increases potential for artifacts, which increases bandwidth.

Brian, I am not sure about the color value / bandwidth relationship but we should be able to test it by getting sheets of colored paper and pointing the camera at each one and measuring bandwidth levels.

Ethan's hypothesis make sense. Derek can you put a few of the Axis and Avigilon images side by side so we can compare. Also, zoom in / sample smaller areas / blow them up - side by side so we can see if there is any measurable difference in quality or rendition that might be driving the bandwidth differences.

If I had to guess, I'd say that different colors are different bandwidths, but it's marginal.

Okay, I picked out a few different spots that I thought held considerable difference between the two cameras.

• The pattern of the macadam is more noticeable in the Q1604, and you can also see old parking lines on the ground clearer in the Q1604. (Wide Intersection)

• The red colored siding in the park scene has different colors throughout, showing the effect the sun has on it stronger than the H3-1MP (Park -- Tight)

• In the next two comparison shots, you can clearly see the severity of which shadows have in the Q1604 scenes, and also a higher level of clarity with the fence, gravel, and grass.

Seems as though the Q1604 will use more bandwidth to capture these details than make them appear "muddier" than the H3-1MP.

That's good analysis.

Which camera has truer color fidelity? :)

e.g., in the parking lot, i am assuming Axis but want to verify with you?

so is this hypothesis now that Axis' truer color fidelity increasing the bit rate compared to a less true camera?

In terms of truer color fidelity, the Axis Q1604 beats the Avigilon H3-1MP hands-down. The Q1604 seems to make the brights "pop" more in the scenes, but couple that with also being able to see those trade-offs in the details, and you can see why the bit rate is higher on the Q1604.

The hypothesis is that a truer color fidelity does increase bit rate compared to a less true camera, but I'm still at a crossroads as to wether or not motion/patterns in a tight scene alter it as much as we may have thought previously.

we have the Axis P3367? the dome 720p lightfinder camera, right? it might be worth testing that in the same scenes, also maybe the Bosch starlight while we are at

i realize this will be more work but i think we have hit on a key point

One thing that's particularly interesting is that at wide FoV the two cameras are close in terms of bit rate - 2.1 avg vs 2.6 avg, it's only at a narrow FoV where presumably the Axis camera can 'see'/'capture' very fine details like the nuances in pavement that the bandwidth jumps - Axis doubles, Avigilon stays the same

I am going to send an email to Axis tonight and cc: you guys

IMPORTANT: Make sure the 2 cameras are aimed at the exact same areas. In a number of your shots, one camera is targeted slightly to the left or right, off by a few feet.

The FoV width is the same but the cameras are angled slightly differently. I don't think this is throwing the test off but it can raise an issue if people want to object (i.e., sure the other camera had a more complex / difficult FoV to monitor because it had the sign in the FoV but the other did not).

Recommend: Next week, the two of you go out and do a field trip with the old setup / tripod / 6 cameras / etc. and get everything perfectly lined up for a formal shootout.

Also, when we dot this field trip with the full kit, we will have subject + license plate to verify that the level of 'real' details captured are the same.

FYI, with a Q1604 pointed at the carpet, bandwidth jumps dramatically as sharpness increases.

• Sharpness 0: ~7.3 Mb/s (also, that's a tremendous amount of bandwidth for a still scene, due to the patterning on the carpet)
• Sharpness 50 (default): ~14 Mb/s
• Sharpness 100: ~24 Mb/s

I never gave a thought to sharpness before this, but I can see overly sharp defaults or people messing with settings have a negative impact on bandwidth, absolutely.

Pictures or it doesn't count!

Seriously though, I'd be curious to see the visual differences. I am very interested in this.

Also, it would make my day to get bandwidth measurement and a snapsot of Avigilon in the same FoV so we can establish another comparison.

Pictures here

I won't embed them, because you need to see them full size to truly see fine differences.

The Avigilon is pointing at the ground right now, so I'll make your day this afternoon.

"Meet Sharpness, the Bandwidth Killer"

Avigilon:

• Sharpness 0: 1.44 Mb/s
• Sharpness 50: 2 Mb/s
• Sharpness 100: 2.16 Mb/s

Images are in the folder. Not as a big a jump, but still an increase. Visually, there is not as much difference in the Avigilon images as in the Axis images, and the Q1604 just sees more detail, period.

Comparison Q1604:

Two questions:

To the human eye, does sharpness 100 more accurately reflect the detail in the carpet?

Does this 'detail' captured actually reflect into material increases in capturing detail meaningful to security users - faces, license plates, etc.?

Sharpness 100 sees more detail than I can see with my eye at the same distance. By far. I'd say I see somewhere between levels 0 and 50. Derek, agree?

And my eyewear prescription is up to date.

I think doing a sharpness test is probably a quick, but potentially extremely useful thing to do.

I took a more academic approach to this ongoing bandwidth analysis, and I found some rather interesting results. I tested color, contrast, sharpness, and brightness at levels 100 and 0, leaving everything else at defaults for each individual test.

• Here is the FoV I used to run this test:

• H31MP: 250 kbps
• Q1604: 703 kbps

As you can see, its a very simple scene, with no motion whatsoever. I left some carpet in on purpose, to see if there was a strong influence during the shaprness portion of the test.

Color

Next, I tested for differences in color and the impact it has on bandwidth. The below screenshots are between color 100 and color 0. I'm still somewhat shocked by the hike in bandwidth the Q1604 has, even with truer color fidelity.

Color 100:

• H31MP: 772 kbps
• Q1604: 1146 kbps

Color 0:

• H31MP: 232 kbps
• Q1604: 567 kbps

Contrast

I was shocked by the difference contrast has on bandwidth. I first ran the Q1604 at contrast 75, since at 100, it looked unusable, but I then ran it at 100, and I saw a huge difference in bandwidth consumption.

• Q1604 contrast 0: 346kbps
• Q1604 contrast 75: 1199 kbps
• Q1604 contrast 100: 2280kbps
• H31MP contrast 0: 180 kbps
• H31MP contrast 100: 489 kbps

These are enormous jumps in bandwidth, and again, I was shocked by what I saw. Again, I'm not too sure someone would run their surveillance cameras at 100 contrast, but this is a solid answer of "NO!".

Key Findings

Sharpness seems to depend on the details in the scene (makes sense), but the difference in sharpness levels between the Q1604 and H31MP are interesting (Especially since there is nothing incredibly pattern heavy in the scene, sans the bit of carpet). At sharpness 100:

• H31MP: 272 kbps
• Q1604: 1175 kbps

And at sharpness 0:

• H31MP: 232 kbps
• Q1604: 534 kbps

Not much difference with the H31MP, but a great difference with the Q1604. Carpet does look more refined in the Q1604 photo in the dropbox, as apposed to the H31MP.

Differences in brightness were rather minor. The Q1604 at brightness 100 was 603 kbps, and at 0 was 577kbps. The H31MP at 100 brightness was 242 kbps, and at 0, 164kbps.

Color and contrast made both cameras variate marginally, with usually a 2x - 3x differene in bandwidth between color or contrast 0 and 100.

I'll throw up a bandwidth spreadsheet soon.

The other photos are uploaded on dropbox.

This is good work. At this point, we need to try with a few other cameras and monitor their bandwidth variances to see how much of a pattern there is across cameras / manufacturers. Please do that.

Bosch / Arecont whoever - make sure fps and resolution set are the same.

How does Sony / Arecont sound?

new Sony with VBR, sure.

also if you have the Axis M114, that would be excellent because then we could contrast Axis across levels (from high end Q1604 to low end M114)

yes, we do have the Axis M114. So, i'll do Sony, Arecont, and the Axis M1144.

Just a quick update, I did not use the Sony (CBR), so I just went with Arecont and M1144.

For tomorrow, when I'm testing the I-Frames, I was thinking of using the Q1604, M1144, Avigilon H3-1MP, and a Bosch 720p camera. Also, what GOP levels would be good to test? I was thinking since the default is 30~32, perhaps 50,60,75,90 and 120.

Would you like me to change any of the color/contrast settings during this test? I'll have 2 different objects, one at a time, then together, to see if a certain motion is more bandwidth consuming than another. Should be interesting tomorrow.

On the I frame test, ignore GOP levels and concentrate on the number of seconds between i frames - try 1, 4 and 10 - 1 second is the most common, 4 seconds is normal 'long' and 10 seconds is real long.

For the I frame test, ignore color/contrast. Just use defaults.

I compiled all of the results from today's indoor bandwidth tests here, and also noted the variance between the baseline measurements and different setting changes.

Very important - when you enter numerical information in spreadsheets keep them in a numerical format. This allows us to run calculations on them, which in turns makes it easy to find patterns.

For instance, instead of '703 kbs', just enter '703' and put the kbs in the header.

Also, for variance / difference, instead of '~1.6x', enter = test bandwidth / cell holding baseline bandwidth number

I can then take those numbers and average them across different grouping, etc.

Also freeze the header top row and first column, it makes it easier to scan across the data sets. I did that here. See the difference it makes.

What's with the M1144? Seems like it's bandwidth is going crazy? Way out of line with the others?

Also, I am confused that some cameras actually increase bandwidth consumption with color min. I mean, it's essentially grayscale so presumably bandwidth should go down?

Across the board, brightness did not seem to make a difference (either high or low). Interesting. I'd be curious to see the pics here to see what the image looks like.

Ok, two other test scenarios to add:

- For the i frame test, do this - Device To Simulate Motion swivel the arm back and forth and see what artifacts / issues you get (at various i frame rates, etc.)

- For the bandwidth / color / saturation / etc. tests, measure the Q1604 with WDR or vs off. I am curious if this has an impact. One of the chip manufacturers has seen this happen with other cameras.

John

I made adjustments to the spreadsheet based on your guidelines, and it does look better with the header row and first column frozen.

The M1144 does a lot of artifacting when color and contrast are at max. The images are below.

I uploaded the M1144 and Arecont screenshots onto dropbox, as well as the recorded video clip of the max color M1144 to see the artifacting.

Working on the toy train measurements now, and I'm also going to run the "tripod swivel" motion test as well.

A major discovery was found while going through some of yesterday's and today's motion tests:

• VLC media player (the software I was using to capture RTSP streams for the bandwidth/motion tests) does not default to using the GPU for decoding. This is very important!!! This change makes the images from the stream look like this:

• To this:

• You can still see a bit of motion blur in the bottom, but that heavy artifacting and doubling are gone. To change this setting, go to Tools > Preferences > Input and Codecs > Check the box that says "Use GPU accelerated decoding".

That's very weird. A single stream is causing such artifacting. Even without the GPU, it would seem a laptop would be able to handle this. Have you checked what the CPU usage was with and without GPU enabled for decoding?