Using Quicktime Player Pro 4.0, export the movie using the "movie to dv stream" option. It should now load in movieplayer on the SGI. (tested with video - don't know if audio does the right thing).
This is an extremely useful combination for video recording, editing, and use for large training sets of video. The program to run is dvlink (man also avc and raw1394). It can control the deck through the 1394 cable. You must be running 6.5.2 or better.
Use Premier (sp?) or moviemaker to create the movie you want. LEAVE ABOUT 10 SECONDS OF BLANK SPACE AT THE BEGINNING OF THE MOVIE IN MOVIEMAKER. This will be important to not lose the first few seconds in the next step. Moviemaker allows you to add audio tracks.
Bring up dvlink. Under Options, under record mode, select video to tape. Hook up the DV deck. Press record on the dv deck and then play on dvlink.
If this doesn't work, moviemaker has a print to video tape option under the File menu. This uses the standard NTSC output jacks. <\li>
To record from the deck to disk, select the appropriate "Record Video" device (usually something like "AVC Device, Phy ID 0"), the proper channel, and give it a file name on the local hard disk ("/tmp/my-dv.dif" for example).
It has to be a file on a local disk running xfs instead of nfs to be fast enough (df will tell you the type).
Press the record icon
IMPORTANT: You can always override the SGI's control of the vcr through the VCR's physical buttons. SOOOO, keep your hands off while recording.
If dvlink gives a "buffer overrun" error, then it couldn't write to disk fast enough and it dropped frames. Make sure no one else is using the local disk and that no other intensive processes are running and try again.
Same procedure as above except this time you must type in timecode values for the "In Point" and "Out Point". Rewind the vcr to a play before the In Point. Press record. The tape will start playing and will continue to play, but the SGI will only record to disk the segment you specified.
In doing computer vision, it is important to have repeatable training and test sets. This SHOULD be the advantage of going digital.
To test how well the GVD900 and SGI did, I (Thad) took some rather crummy video already recorded on a DV tape several weeks ago. It had been played several times already. I specified an In Point at 00:00:10:00 and Out Point of 00:00:15:00 for a duration of 5 seconds plus one frame.
I recorded the video to disk 10 times and did a bdiff between the first recording and the rest. Any time dvlink complained of a buffer overrun I re-did the trial (with good reason - the video is significantly changed when this happens).
On average 1.5 bytes of the 18,120,000 byte file were different. I suspect that the byte errors were actually due to a single bit error but did not confirm this. Making this assumption, I had between a 10^{-9} to 10^{-8} bit error rate (or, in other words, a bit every 3.33 seconds of video). This is quite excellent compared to my previous experiences with the reproduceability of video stored on Hi-8 or BetacamSP.
So, in conclusion, while DV is about a 5-7.5 to 1 compression when compared to raw, uncompressed 640x480 24 bit video and doesn't devote as many bits to color as I'd like (its a 4:1:1 scheme), the reproduceability more than makes up for it if your purpose is evaluating your vision system.
Note however that in real life camera noise will still be a factor, so make sure your training and test sets span the space of possible noise characteristics for the application. Thad Starner
CCG Home Page
Last modified: Thu Jul 29 15:03:33 EDT 1999