MAN WITH A MOVIE CAMERA (MOTION TYPE DATA REDONE) (1929, Soviet Union)
directed by: Dziga Vertov

BACK TO THE DATABASE

IMDB link: http://www.imdb.com/title/tt0019760

Submitted by Adelheid Heftberger, Yuri Tsivian, Barbara Wurm on 2008-06-12

Adelheid Heftberger, Yuri Tsivian, Barbara Wurm's comment:
Adelheid's motion type data in advanced mode as follows:

NM: No motion
FF: Freeze Frames
SMC: Slow Motion (Camera)
NormalN: Normal (naturally)
FastN: Fast (naturally)
FastC: Fast Motion (Camera)
BF: Black Frames, Intertitles, 1-Frames

This submission is one of a sequence. We (the three names above) plan to go on submitting timing data for Man with a Movie Camera and segments thereof in both simple and various advanced modes, and discuss the results in comment boxes as we go. This look-and-talk sequence is part of a project "Digital Formalism: The Vienna Vertov Collection" (www.digitalformalism.org) made possible through a cooperation of three institutions: the Department for Theatre, Film and Media Studies (Vienna University), the ?sterreichisches Filmmuseum, and the Interactive Media Systems Group (Vienna University of Technology).

This measurement is a frame-by-frame count. The count is based on a 35 mm print preserved in the Vienna collection (provenance: Gosfilmofond of Russia) digitized and annotated using Anvil software. Adelheid Heftberger was in charge of annotating this digital copy and locating 5 splits between the film’s 6 reels (not yet identified in this submission). The result was proofread and checked by Heftberger and Edith Schlemmer against the 35mm print on a Moviola (Arri) viewing table and by Heftberger and Tsivian against a shot list compiled in the 1980s by Vlada Petric and Roberta Reeder using a 16 mm print from the Harvard Film Archive.

To translate frame numbers into seconds we proceeded from an assumption that the correct projection speed for this film was 24 fps. This assumption (which to some may sound pretty unorthodox) was based on two kinds of evidence. One is Kevin Brownlow’s detailed investigation into projection practices in the late 1920s which those interested will find at http://www.cinemaweb.com/silentfilm/bookshelf/18_kb_2.htm. The other is a document from the Vertov archive – a music sheet complete with time markers used by an orchestra at the film’s opening in 1929 (see in: Yuri Tsivian, "Dziga Vertov's Frozen Music: Cue Sheets and a Music Scenario for The Man with the Movie Camera," Griffithiana, No 54, October 1995, pp. 92-121) from which it follows it was then shown at 24.


Name:
NM
FF
SMC
NormalN
FastN
FastC
BF
Number of shots:
207
13
32
624
465
273
115
Length(min):
8.1
0.87
2.25
34.98
12.16
4.8
3.25
ASL(sec):
2.3
4
4.2
3.4
1.6
1.1
1.7
MSL 2.2 4 3.7 2.8 0.9 0.4 0
MSL/ASL
0.94
1
0.88
0.83
0.57
0.38
0
StDev 1.6 2.6 2.5 2.8 1.9 1.4 3.1
Min 0.3 1.2 1.1 0.1 0.1 0.1 0
Max 8.5 10.8 12.6 22.6 16.6 7 17.1
CV 0.7 0.66 0.59 0.85 1.19 1.37 1.84
Display?
Color              
Loading...

Step: Vertical resolution: Height:
Degree of the trendline: Moving average : Color code?


Users' comments:

Author: Heidi to Yuri and Gunars Date: 2008-10-30

Thank you Gunars, this is like early Christmas ;-)

Before we start the discussion, let me remind you of something Yuri said when we submitted the first graph and which can help to analyse the result of our submission:

YURI: Our hypothesis was that the two variables "covary positively," namely, that the cutting rate would increase with the intensity of movement within shots. We will first need to come up with two scales, a speed-scale for in-shot movement and a simple ASL scale for the cutting rate, and, by plotting them on an Y- and X-axes, we will evaluate the degree of the correlation we are looking for. There are graphs and formulae in statistics which allow to establish whether this or that correlation exists and how strong it is.

Here is a first diagram, is this what you thought of?

One observation already: We can see that at the extremes we have the Camera Motions: Slow Motion with the highest ASL and Fast Motion with the lowest.

 



Author: Yuri Tsivian Date: 2008-11-01

Good work, Heidi, Gunars, but there are two or three things you need to correct. 

1) Gunars: the legend you included in the "author comments" box is incorrect: these are the FIRST MOTION TYPE SUBMISSION CATEGORIES, but this time we decided to collapse the "SMN: Slow Motion (naturally)" and "NormalN: Normal (naturally)" into one category, NormaIN. In your submission and in Heidi's comment there is no SMN any more, but in the lkegent there is.

2) Heidi:

       a) I don't think we need to take the BF category into account, for we dumped thing into it which we deemed irrelevant for our investigation, remember?

        b) A better (if not the only) way to represent correlations graphically are scatterplots, they say, not linear graphs. Scatterplots Gunars tells me, can be generated within excel files, can you figure out how, or should Gunars help?

         c)   In your graph, ASL are arranged from 0 to 5 in ascending order, exactly as your old maitre René Descartes taught us to do. But remember: on Cinemterics we bravely go against Cartesian principles (timid Barbara deserted us for that). What we are inteseted in is how high are cutting rates, not in how long is this or that shot. So, to the Y axist should be turned upside down, preposterous as this may at first look.

          d) We have a strong case, I agree, but the NM category clearly jumps out of the trend. Remember we discussed a possible reason once already? So, let us spend some time thinking about this, looking back at the film (you in your Anvil, I in my mind). In other words: why do they cast still shots faster than they do "normal movement shots"?

 



Author: Gunars Civjans Date: 2008-11-04

I fixed the legend.

Scatterplot won't work in this case, because you don't have a precise value for movement (x-axis), just a few integers, so it will look like this:



Author: Yuri Tsivian Date: 2008-11-04

Doesn't tell us much, does it?

Is it the spreadsheat that cannot plot the scattergraph, or what? I looked up my Teach Yourself Statistics  on paired data and best-fit lines, and it does not seem to have a problem with as few integers as you wish. See for instance:

I am afraid there was something wrong with our thinking here, Heidi, dear.

Look at your graph again, my child, and let us figure out what is wrong with it. I mean the graph you posted on October 30, not Gunar's blueish rectangle right above us.

Let's forget for a moment what I wrote a few days ago about reversing the Y-axis. This is easy to do, the matter of techniques. We'll do this later on.

The important part are conceptual things. First of all, remove, in your mind's eye, the line that you used to connect the dots. The line is misleading. It makes the dots look as if they were part of a process, of an economy's rises and falls, for instance, while in reality the linesignifies nothing, for both the Y- and the X- axes are DEGREES not STAGES.

Done? Good. Now instead of a curve you have a conglomeration of dots, right, Heidi? The next question is, what are these dots and what they tell us.

The first thing we see about these dots is, they are scattered. In fact, this is already a scattergraph. The first thing we want to know about the dots, is how widely they are scattered. This is what scattergrams tell us. If the dots are evenly scattered throughout,  there is no correlation; it they are aligned, there is a perfect correlation. If they are scattered but not widely, if they tend to group around or along an imaginary line, much like the Milky Way does, there is a correlation, weak or strong, depending on how tightlyknit a pattern the dots form.

To establish what the prevailing tendency is, statisticians use what they call the "best-fit line." Not a line that connects the dots, but a line that cuts through their array, like theone on Figure 11.7 in the textbook example I pasted above.

Two last things before you do it. First. Of course, we do not need the last dot on the right, for early on, we used this category to dump all irrelevant shots in it. Secondly, don't forget to reverse the damn Y-axis, for our ultimate value is SPEED, not slowness. If you are worried about the ascending order, use minus-ASL, perhaps.

And we still need to discuss the No-Motion shots, but later.



Author: Yuri Tsivian Date: 2008-11-05

On second thoughts, Heidi, let's better wait before we go on with those graphs. I just took a bicycle ride along lake Michigan and as I pedalled on it occurred to me that what I have written above may be, how to put it, incorrect. After all the data you obtained --

2.3
for No movement,
4
for Freeze Frame
4.2
for SloMoCam
3.4
for NormMo
1.6
for FastMoNat
1.1
for AccelMo

-- these six pieces of meaningful numbers are, after all, averages for six different categories.  What is normally taken into account in scattergraphs are not averages, but actual measurements projected on a quasi-continuous grid. For instance, if there were ten people whose pulse was measures there should be ten crosses scattered across the graph. In our case, for this to be a scattergraph in the canonical sense of the term there should be 1729 dots, not 6 -- the number od actual shots in the film (no, less -- for we would exclude 115 shots that went to BF, the irrelevant category, but still, quite a number). Now the reason why they formed the dense milky way that I was talking about (as they should have on the blue graph which Gunars submitted last night) is that the X-axis as it stands now is not a continuum, but a scaled line, with values assigned to 6 notched on the line, not to segments on it. As a result, instead of scatter we have  7 thin fountains that appear to be shooting up with different degrees of intensity. 

Now, my question is: can we perhaps trick the system into giving us a scattergraph proper without, of course, falsifying the picture we are seeking to draw up? What if we divide the X-Axis into 6 segments: A to B = segment 1 for No movement; B to C = segment 2 for Freeze Frame etc., till  the end of the scale? And somehow tell the system to randomize the shot data within each segment?

 



Author: Gunars Civjans Date: 2008-11-05

Randomizing the speed value might not be mathematically sound but it gives a better view of shot numbers in each category. It was difficult to see the individual shots in the "fountain" scatterplot. Below are the two graphs with Y-axes reversed. First one with x-values randomly distributed to take up an area from the whole number to +-0.5. There is still no overlap and you can still distinguish the fountains, only they are much thicker now. Second one without the randomization. The correlation for both graphs is about 0.2, which is considered small (correlations range from 0 = no correlation to 1 = perfect correlation). If you remove the X=1 value (No movement) then the correlation increases to about 0.4. Black frames skipped.



Author: Yuri Tsivian Date: 2008-11-06

This about sums it up. Having watched me tinker with self-taught statistics, Gunars has taken pity and, like deus ex machina, fixed it all. 

The bottom line: there is such a thing as the Heftberger Correlation. Distinct from  what I initially thought, it is not strong and it is not linear. It becomes stronger as we move left to right from the static  side of the scattergraph toward its more dynamic wing. 

As you may recall, Heidi, I suspected from the outset that we would have problems with the first of our categories, the Non-Motion shots. The thing is, non-motions shots are not only about no motion. They take part in too many different games, serve too many masters, are used for too many purposes to be able to tell us much about the speed-driven correlation. Take, for instance, the shot-reverse-shot editing pattern, like when a young lady practices sharp-shooting. We see her, aiming, and then the target, her, the target, her, the target etc -- and then she pulls the trigger and the target is set in motion. This is a fast-cut sequence, it builds up tension, yet technically it largely consists of non-motion shots (much like when two motionless cowboys are having a showdown in a Sergio Leone Western). There are more than one case like this in MWMC, as the one with the mannekens "looking" at empty streets. The looking-looked-at pattern, in order for it to cohere, must be a fast one, and it is, yet technically it consists of still shots. What we have here is a effect of interference: the Conventional Editing Factor interferes with the Event-Driven Factor and the purity of our statistical data is therefore compromised.

One could go through MWMC once again and cull out all suchlike shots and resubmit again, and it's up to you, Heidi, if you want to do so (after all, it's your Correlation), but I think we now know enough to go on.

What do we know then, aside from the mere fact that there is that Hefberger Correlation (HC)?

Three things.

1) That HC is weak (0.2) when we factor in the polluted category, the Non-Motion shots.

2) That HC becomes stronger (0.4) if we take the non-motion shots out of the equation:

 

3) And if we decide to take out of the equation all shots with no motion in them (that is, two categories, No Motion and Freeze Frames) we will obtain what we can name the Particular Heftberger Correlation (PHC, valid for moving shots only) which will be even stronger, and, which is important, linear:

 



Author: Yuri Tsivian Date: 2008-11-09

Of course, it should be added, the HC effect may only work with MWMC; to see if it works with other films or even with other documentaries of the kind similar measurements must be conducted.

This means we need more volunteers. People interested in following this conversation from where it started are encouraged to go to a string of comments under Man with a Movie Camera (Motion type data): (6) ASL 2.3 and Man with a Movie Camera (Chelovek s kinoapparatom): (6) ASL 2.



Author: Heidi to Yuri and Gunars Date: 2008-11-11

Wow. One week absence and then see a high-level discussion going on around a (casually done) graph in excel thrown into cinemetrics.

This first try was actually really only attempted to give an impression (visual) of the average shot length of our categories. And of course the line between them doesn't mean anything - only to confuse the viewer, of course. But: Still seemed to be a good trigger for the comments above and made at least Yuri pedalling around lake Michigan. That leaved me satisfied. As for the graphs, I still would like to work something out a bit easier to understand at first glance, also a kind of table that can be used for future films. Or do you have that already Gunars? And for your graph you used the single shots with their category number, right? Should have done that in the first place of course.

So Yuri, would you leave the NM category our for further submissions? I think it should be done, the same with BF (Irrelevant - which we should rename anyway, because BF derives still from "Black Frames") - later on data can always be "left out". I have some more Vertov films at hand, which could always be done. But we can come back to that question (or further projects) whenever we want.

 



Author: Yuri Tsivian Date: 2008-11-22

This, dear Heidi, really depends on the film you submit. We cannot simply say, I am afraid, that the HC will be applicable for all types of movies in the future, and have a single immutable approach to all films. The HC is very likely to work with most of the "city symphony" types of movies and with documentaries of the "11th year" type, for in both speed is thematised and foregrounded; but how it might work with more "normal" films is an open question. With action films like "Speed" it might; but what about the drama genre, for instance?

And remember, my angel, that MWMC is a film that has seven speeds, much like I fancy professional race-cars might.  How many films can we name that can, in the span of an hour, change from freeze to still, to slow-mo, to normal, to normally fast and to accelerated fast? Off the top of my head I can mention only Ou Ning's "city symphony" film "San Yuan Li" (2003, 44 minutes), the most extraordinary modern variation of the Vertov-type city symphony I have seen. Maybe one day you try your hand at measuring this Chinese film? The Heftberger Correlation will probably work for it.

In other words, with most "normal" films you'll have to deal with only two or three categories: stop, slow, fast. How will it work in those cases?  No idea. Would I like to know? Very much so.



Author: Heidi Date: 2008-11-22

I will have to take some lessons in Statistics, clearly. Also my sister (as a biologist) is not bad at it and can explain some things maybe.




Add your own comment:
Your name:




Enter code from above picture: