Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!sdd.hp.com!usc!apple!agate!ucbvax!ADS.COM!Vision-List-Request From: Vision-List-Request@ADS.COM (Vision-List moderator Phil Kahn) Newsgroups: comp.ai.vision Subject: Vision-List delayed redistribution Message-ID: <9011060500.AA00748@deimos.ads.com> Date: 5 Nov 90 17:59:09 GMT Sender: daemon@ucbvax.BERKELEY.EDU Reply-To: Vision-List@ADS.COM Distribution: inet Organization: The Internet Lines: 181 Approved: vision-list@ads.com Vision-List Digest Mon Nov 05 09:59:09 PDT 90 - Send submissions to Vision-List@ADS.COM - Send requests for list membership to Vision-List-Request@ADS.COM Today's Topics: RE: Optical Flow in Realtime Standard (or at least famous) pictures - where to find them SGI and ground-truth for shape from motion algorithm Fingerprint ID Abstract of Talk on Computer Vision and Humility ---------------------------------------------------------------------- Date: Fri, 2 Nov 90 23:33:13 GMT From: Han.Wang@prg.oxford.ac.uk Subject: RE: Optical Flow in Realtime > (1) Is there any company or research lab which could >compute on grey images (256x256) image flow in real time? I have achieved the rate of 2~4 seconds in computing optic flow using 8 T800 transputers on 128x128 images. This is only along edge contours (Canny). >We at BMW are developing a lateral and longitudinal >controlled car, which should (for experiments) drive >automatically and which might be in the future an intelligent >assistent to the driver, in which form soever. > >We will use (if available) this techniques to detect >obstacles, that are lying or driving on the street, In oxford, we are building a system of bybrid architecture using Sparc station, Transputer array and Datacube to compute a 3D vision system DROID (Roke Manor Research) in real time which can effectively detect obstacles in an unconstraint 3D space. This is however not based on optic flow. It uses corner matching instead. So far, we have succeed in testing many sequences including a video camera carried by a robot vehicle. This experiment will be demonstrated in Brussels during the ESPRIT conference (9th - 16th Nov. 1990). regards Han ------------------------------ Date: Fri, 2 Nov 90 11:40:06 EST From: John Robinson Subject: Standard (or at least famous) pictures - where to find them We are searching for raster versions of "famous" monochrome and colour images. Actually, any format will do if we can also get access to a format to raster convertor. We are particularly interested in getting hold of: Girl and toys, Boy and toys, Gold hill (steep village street with countryside) Boilerhouse (picture with lots of shadows), Side view of man with camera on a tripod (actually there are at least two pictures of that description around - we'd prefer the one with the overcoat), The various portraits from the 60s of one or two people that are often used, Any single frames taken from videoconference test sequences. Anything else that fulfils the following would be appropriate: Good dynamic range, Low noise, No restrictions on copyright, Portraits completely devoid of sexist overtones (e.g. not Lena), Is there an FTP site with a good selection of these? Thanks in anticipation John Robinson john@watnow.UWaterloo.ca [ The Vision List Archives are on the build. Currently, of static imagery, they contain Lenna (girl with hat) and mandrill. A collection of motion imagery built for the upcoming Motion Workshop (including densely sampled and stereomotion imagery) is also in the FTP accessible archive. If you have imagery which may be of interest and may be distributed to the general vision community, please let me know at vision-list-request@ads.com. phil... ] ------------------------------ Date: Thu, 01 Nov 90 19:38:35 IST From: AER6101%TECHNION.BITNET@CUNYVM.CUNY.EDU Organization: TECHNION - I.I.T., AEROSPACE ENG. Subject: SGI and ground-truth for shape from motion algorithm I am presently working with 3-D scene reconstruction from a sequence of images. The method I am using is based on corner matching between a pair of consecutive images. The output is the estimated depth at the corner pixels. The images are rendered by a perspective projection of 3-D blocks whose vertices are supplied by me as input to the program. However, the detected corners are not necessarily close to those vertices. In order to obtain a measurement of the accuracy of the algorithm I am using, the actual depth at that pixel is needed and I tried to recover it from the z-buffer. I thought that (my station is a SilliconGraphics 4D-GT) the z-buffer values (between 0 and 0x7fffff) were linearly mapped to the world z-coordinates between the closest and farthest planes used in the perspective projection procedure available in the Sillicon's graphic library. The results however don't match the above hypothesis. I tested the values of the z-buffer obtained when viewing a plane at a known depth and it was clear that the relation was not linear. Can someone enlighten me about how the z-buffer values are related to actual depth? I know there is a clipping transformation that transforms the perspective pyramid into a -1