![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
![[identity profile]](https://www.dreamwidth.org/img/silk/identity/openid.png){kind=small}
While hunting down the first major reference to the problem of the scoring of segementations, I was led to this paper published the year I was born:
W.A. Yasnoff, J.K. Mui, and J.W. Bacus. Error measures for scene segmentation. Pattern Recognition, 9(4):217-231, 1977.
It's a pretty good paper, and it covers some techniques which, while basic and obviously flawed, are still used since no one has really come up with anything better. It helpfully includes the full expansion for the term "pixel" in case you weren't familiar with it. (Picture element.) In order to score output from their system, they needed a set of correct segmentations to compare them against. (We'd now call this the hand-labeled ground truth.) But they were still at the beginning of interactive graphics and couldn't just paint different regions of the image using a mouse. So they had to be creative. They took the pictures to be segmented and project them as 35mm slides directly onto a monitor. Then using a text editor and joystick, they filled in the different regions with different characters. That was the basis for the rest of the system, too. All their sample output is ASCII art.
This led me to a reference to an even older paper, which seems to be the first paper talking about the segmentation problem ever. I stopped by the library to pick up the book it was in, since I couldn't resist having a 1968 reference in my thesis.
J.L. Muerle and D.C. Allen. Experimental evaluation of techniques for automatic segmentation of objects in a complex scene. In Pictorial Pattern Recognition: Proceedings, page 3. Thompson Book Co., 1968.
I greedily skimmed the entire book on the bus ride home. It was like reading about alchemy or looking at the Wright Flyer. Just... adorable. In the 28 articles inside, there isn't a single use of "pixel" that I could find. Picture element, image element, matrix component, cell... all the papers are talking about pixels, but no one has decided what to call them yet. In one paper, they warn that when scanning, if you use the wrong cell size the object can end up 'fragmented'. After blinking at this, I realized that they meant 'pixelated'. Lots more ASCII art output, too. Even for real images, printing them out on a teletype was the best they could do. One paper had a screenshot of an oscilloscope level CRT, used for a kind of rough vector output.
The segmentation paper itself was dealing with images scanned from transparent slides. The scanner they describe worked a pixel at a time, shining a light through to a sensor then moving on. 64 levels of grey -- 6 bits. They couldn't load the entire image at once, due to memory constraints, so they'd read in a pixel, compare it with neighbors, then store the yes/no results in a much smaller array in memory. When all this was done, they then used the same scanner for output. The slide was replaced with a sheet of photographic paper. The scanner was then used to burn each pixel onto the paper, which was then taken to a dark room and developed. Utterly crazy. And not a little bit humbling, that they were doing so much with so little.
ETA: It has been pointed out to me that I've probably met the author of the first paper, and am definitely only a single Erdos-number-step away from him. Small world.
W.A. Yasnoff, J.K. Mui, and J.W. Bacus. Error measures for scene segmentation. Pattern Recognition, 9(4):217-231, 1977.
It's a pretty good paper, and it covers some techniques which, while basic and obviously flawed, are still used since no one has really come up with anything better. It helpfully includes the full expansion for the term "pixel" in case you weren't familiar with it. (Picture element.) In order to score output from their system, they needed a set of correct segmentations to compare them against. (We'd now call this the hand-labeled ground truth.) But they were still at the beginning of interactive graphics and couldn't just paint different regions of the image using a mouse. So they had to be creative. They took the pictures to be segmented and project them as 35mm slides directly onto a monitor. Then using a text editor and joystick, they filled in the different regions with different characters. That was the basis for the rest of the system, too. All their sample output is ASCII art.
This led me to a reference to an even older paper, which seems to be the first paper talking about the segmentation problem ever. I stopped by the library to pick up the book it was in, since I couldn't resist having a 1968 reference in my thesis.
J.L. Muerle and D.C. Allen. Experimental evaluation of techniques for automatic segmentation of objects in a complex scene. In Pictorial Pattern Recognition: Proceedings, page 3. Thompson Book Co., 1968.
I greedily skimmed the entire book on the bus ride home. It was like reading about alchemy or looking at the Wright Flyer. Just... adorable. In the 28 articles inside, there isn't a single use of "pixel" that I could find. Picture element, image element, matrix component, cell... all the papers are talking about pixels, but no one has decided what to call them yet. In one paper, they warn that when scanning, if you use the wrong cell size the object can end up 'fragmented'. After blinking at this, I realized that they meant 'pixelated'. Lots more ASCII art output, too. Even for real images, printing them out on a teletype was the best they could do. One paper had a screenshot of an oscilloscope level CRT, used for a kind of rough vector output.
The segmentation paper itself was dealing with images scanned from transparent slides. The scanner they describe worked a pixel at a time, shining a light through to a sensor then moving on. 64 levels of grey -- 6 bits. They couldn't load the entire image at once, due to memory constraints, so they'd read in a pixel, compare it with neighbors, then store the yes/no results in a much smaller array in memory. When all this was done, they then used the same scanner for output. The slide was replaced with a sheet of photographic paper. The scanner was then used to burn each pixel onto the paper, which was then taken to a dark room and developed. Utterly crazy. And not a little bit humbling, that they were doing so much with so little.
ETA: It has been pointed out to me that I've probably met the author of the first paper, and am definitely only a single Erdos-number-step away from him. Small world.
no subject
no subject
no subject
no subject