Well, Todor, these are pictures of chairs and *you* have little problem recognising them – (agreed, the odd one may cause confusion – but that is a fundamental part of the business of object recognition).
And you can recognise sensory images of such chairs from a distance, with v. limited “3d/solid” information (it sounds like you’re talking mainly about viewing objects as “solids” close up (in the flesh). We detect a lot of “3d” info, you see,. from 2d pictures. If you are suggesting that such object recognition processes in humans and animals are based on “solid” experience of comparable objects, I would agree. But then you will have to give some indication of how the methods you outline apply to robots interacting with solid objects – I wasn’t aware that you were/are doing that. Are you? Ultimately you will have much the same problems dealing with solid objects as with their pictorial images. They too are as formally diverse as the images. Then certainly you/a robot can interact physically with the objects and ascertain if you can sit on them. But there will still be problems of classification – for example, separating stools/walls/boxes and other sittable objects from chairs. And more or less each solid chair in the pictured range still represents an extraordinary **transformation** of the last, You’re not really advancing our, i.e. the general, analysis of the problem, by implying it all can be easily solved. – by geometric operations. You have provided no evidence that they apply to those images, and by extension to their objects – and can explain the uniformities underlying that vast range of transformations. I contend as you know, that it is only by comparing them with the multiform transformations of irregular, more or less fluid objects ranging from blobs and waterdrops all the way to rocks and wood chunks, that we can understand the design transformations of chairs - and not by comparing them with the geometric, uniform transformations of geometric objects. I would suggest it is far better for you to recognise the difficulties of the problem – which you have effectively done by pointing out the enormous visual ambiguities of the chair images – with wh. I broadly agree. Then, I suggest, you have more to offer – it’s useful for example to bring in, as you have done, how the recognition of objects depends on how we physically interact with those objects – whether in this instance we can sit on them. (This exchange with you & Aaron has been valuable for me in underlining how probably essential that dimension is for object recognition). But unless I missed it, your proposals don’t call for simulated interaction of the robot’s/viewer’s body with the objects viewed – such simulation is certainly necessary for us to recognise those chairs- and there is massive evidence for it. (Is anybody in AGI or robotics attempting such simulation?) . From: Todor Arnaudov Sent: Sunday, November 04, 2012 9:17 PM To: AGI Subject: Re: [agi] The Fundamental Misunderstanding in AGI [was Superficiality] Mike, these are not chairs, and my method requires 3D - the *planes* require 3D to be defined. In 2D there are only lines, (closed) curves, (closed) contours (you call them "blobs") etc., and this is a 2D picture, or a set of 2D "objects" (identified/separated by proximity/and low/high contrast), I already told that. The visuo-spatial concept for "chair" is derived from 3D, where the observer can play with the coordinates and orientation of the objects and the "camera"; in a space with gravity, with known (perceivable) dimensions/sizes/distances/strengths of the materials etc., and known sizes of the agents/their parts that would use those chairs, and to which the chairs should match, such as: - average length of the thigh - average length of the leg from the knee to the foot - average length of the back (of a human) We reconstruct 3D from 2D by reverse-perspective and reverse light transforms, and because the G-force and center of mass recognition (that's one way to recognize intuitively where's "down"), and already seen 3D things which project in 2D like given 2D images etc. (some matches are recognized directly, like the legs, some ranges of proportions). The recognized archetypal cases, their legs (gravity), relative dimensions etc. also suggest that the strange ones "might be chairs". However 3D from 2D is an ambiguous operation, it cannot always work correctly (the reason for some of the optical illusions), and that goes even for shaded or color images. Obviously generalizing a class of 3D-objects from a poor 1-bit 2D silhouette projection from one point of view is vastly ambiguous. I already told about that - using vague digested features that cannot be generalized as expected, because the general in the originals is swept away in this representation. ... Also, the recognition of given objects as "chairs" (and not instances of another class) is due to the fact that the other known things/samples are *different* enough. In simplified representations and with a whole lost dimension those differences get reduced, and different classes now look the same (that's one aspect of resolution reduction I talk about). What if I see it in 2D, as it is? You and the context (the most "archetypal" cases) suggest that perhaps the rest are also part of the same class - you suggest that, I also already explained that, but... If any of those pictures is shown separately out of the context and you ask a random guy to tell you what he sees in one word - I don't think he would always say a "chair"... Let's see the figure that is 3-rd on the right, second row from the bottom - nobody would say this is a chair if seen that way in 2D in 1-bit color separately from the context. This is rather a satellite receiver, a deep dish, a cone with a "half-moon" on it. The one on the left looks like a head put on a pillar - it has an ear, a nose, a chin. The one next on the left looks like a barbeque or like something to put a baby in. http://pristine-home.blogspot.com/2012/04/barbecue-time.html http://www.diytrade.com/china/pd/6599510/Electrical_Baby_Swing_BSE900.html The one below looks like it's made from balloons. There's a "magnifying glass", above the "head on a pillar": http://www.mallcarts.com/vdir/vitem/AM-AC099A There's also another satellite - the right-most in the first top row. The figure that is two positions right looks like an (alien) insect-cyclops. The fourth on the bottom row is like a face of a cartoon villain who's spitting something - a spider's web or a puke? The right-most on the top-row is like an alien ray-gun, an antenna, a drill, a brush - not a chair, because from that POV in 2D you can't tell whether the short lines are flat, deep and rectangular, or circular. The one on the right looks like the Alien, or like two burning torches, put on rest "on the thing in the middle" - possibly an alien holding those torches. Etc.etc. - see more suggestions on the picture. If you got a 3D view, the size, the materials of which those are made, their strengths, their transparency, texture etc. then you will got more clues to decide are those really "chairs" or are the other ambiguous objects in the 2D-silhouettes, are toy chairs, small chairs, big chairs, designer chairs or sculptures. The method with planes applies for their most general visuo-spatial structure, and I repeat: it's for abstractions, for words. To specify everything, many words and lower level data are required, and the abstract classes get overlapped, I've discussed that and the raw data in the initial "chair thread". ... >Now show us how any of your methods apply - how they can identify common >blocks/elements of these chairs/ chair images. >You don’t necessarily have to do the whole chair – just show us how your >methods will *begin* to apply to parts. Todor: You're challenging a method which talks about planes and 3D, using a 2D picture having one plane with everything laying on it. In one plane you cannot fit a plane, a perpendicular vector and another perpendicular plane, parallel to the vector. >You are doing here in micro what AGI systembuilders incl. B&B have done in >macro – spent a lot of time detailing a system, >without ever taking the trouble to see whether the method/system actually >applies to the problem – when in fact you should all >have *begun* with the problem. Todor: Sorry, a bullshit, I have visual imagination (tautology), it seems not everybody does. The visual (geometrical) generalization is essentially linear algebra, topology, calculus, and it's obvious. It also emerges from the raw data by simple matching and resolution variations. Well, too much talking - I'm off to work soon. >All your time is wasted unless you can show your efforts really apply to the >chairs - and can actually identify similar >elements in different chairs. Todor: I can identify similar planes in a 2D picture, I can identify similar curves, similar sizes of them, similar topology, similar curvature etc. That's calculus and topology (e.g. graph theory). I already wrote about those parts in maths. >I don’t think – and, I would say, the great majority of philosophers and >psychologists – don’t think you will. Todor: I'm probably partially repeating myself, but most philosophers and psychologists (or I assume yourself, too) don't have 1/15 or 1/20 of my expertise or my talents, or of the talents of other polymaths and universal artists... And yet they believe that they understand general intelligence or creativity better than them... I've done original philosophy in real time, on the spot, as a teenager, "abstract" ethical and whatever problems are trivial to me - they don't understand in formal terms serious mathematics or physics even in their 70-ies, nor can deal with programming/CS/computer graphics (a close friend of the above, which have many sub-branches) - for most people programming is absolutely impossible to grasp even for elementary tasks, the more complex tasks require even more complex mathematics and too big working-memory. On the other hand for ones who deal with it, or for me it's "like ringing the bell" right from when I first started - trivial. They may know some neuroscience, but not for sure, and they cannot make the associations to mathematics and CS, they're "mathematically blind". Also perhaps blind in terms of physics. They probably lack serious visual processing capabilities, like imagination of an artist and a movie maker - to see movies of the things and how they can be transformed or built in a sequence, and to see the outcome of many transformations immediately, without having to do them in reality. They probably lack the visual and physical processing of an artist or engineer in terms of easily playing with perspective, 3D, mechanics (support, center of mass, balance, statics, kinematics, dynamics, projection, dimensions/sizes...). If so - yeah, probably they cannot see what I'm talking about, everything shows that they lack a lot of cognitive faculties that I have. * Todor "Tosh" Arnaudov * -- Twenkid Research: http://research.twenkid.com -- Self-Improving General Intelligence Conference: http://artificial-mind.blogspot.com/2012/07/news-sigi-2012-1-first-sigi-agi.html -- Todor Arnaudov's Researches Blog: http://artificial-mind.blogspot.com On Sun, Nov 4, 2012 at 1:23 AM, Todor Arnaudov <[email protected]> wrote: "Todor: - Identify the gravity force, their direction - Identify a support plane/surface - Identify a plane that's perpendicular to the gravity vector and and has a support - Size must be appropriate for sitting (area/length of the planes of the ass, of the chair) (Some additional) - The center of gravity must be low enough so you don't fall after sitting - The sitting plane has to be reachable, and must be big (or small) enough There you go." Mike:> Well, at least you’re trying. Todor: See also the old explanations, I'm lazy to search and copy. >This isn’t really coherent – it sounds like you’re saying that all these chairs must have a seat/”sitting >plane”/”support plane” to classify as chairs. >But what does a support plane or sitting plane look like? Draw it. There are a lot of support planes >on these objects, just as gravity is acting on them at many different points – and there isn’t one >concentrated direction of “the gravity force.” Todor: "A plane" is a mathematical concept. It looks what a "plane" look, it's not a drawing, it's a correlation of points in space, it can be in voxels, a volumetric integral (if the thickness is taken), or surface integral which if seen in low resolution is just a plane. If it doesn't turn into two planes or if there's no plane after resolution reduction - that would be something else. There are not a lot of support planes, this is the lowest resolution representation, and there is a concentrated direction - it's an "averaged vector" and goes through the center of mass, which usually is through the center of the plane that is perpendicular to the "ground", the ground is found by the direction objects fall if there's no support etc. (Center of mass is what allows drawing realistically human figure in action, and it is one of the ways to anticipate motion in picture and that somebody is going to fall - when she's out of balance, which is determined easily for the human body, apparently by our experience of watching humans moving and falling). >If you mean that such a plane is any that a bum can sit on – what does the bum (or human-figure- >with-bum) you are going to apply to these drawings look like, and where are you going to apply it? >How do you know where to apply it – where and how to sit on these chairs?. >And if your sole criterion of a chair is a seat/sitting plane, how are you going to distinguish chairs >from swings? Todor: I have already written about that and answered some of those question. Yes, there are many criteria, including the materials, here I'm talking about the visuo-spatial static domain (voxels/contours/planes) and gravity ("statics" in physical terms). There are many specific details, which are not essential. A "swing" is obvious - it can swing and naturally swing ("swinging" is the trajectories which are called so) - it's not static, it's dynamical, i.e. the coordinates of those planes are supposed to change. Yes - they can be changed for chairs with wheels, and one can swing on any chair (but may fall back, too), but the directions are different. In visuo-spatial low-resolution domain, they are similar, the details, materials, static or dynamic, the axis where it rotates or swings, the location of the chair/swing in global coordinates (the context) etc. etc. of course add to the definitions, and all can be classified in sub-classes for specific types of chairs, swings or whatever. I am talking about general cases, in low resolution, otherwise there couldn't be words, but just exact pictures, not concepts. I've already discussed precisely this issue for chairs and buildings. >Your method doesn’t add up to a coherent form – a “uniform block” - that all these chairs have in >common – It's the opposite. > or a coherent method of sitting on them, and identifying them that way. Todor: The "method" is to compare the size/surface of the bottom to the surface chair, the height to the ground, the strength of the support, of the back etc., that defines whether you can sit and will you fall. Those comparisons are literary comparisons: matching the size/area/height, numbers. There are many ways of "sitting" etc., but "stool" or "chair" are *general*, these are words - their purpose is to tell as much as possible for the context with as little data as possible. >P.S. And what happens if you’re confronted with a chair upside down, or whose seat has been >smashed? We can still recognize a chair if the seat is missing. How? Todor: Upside-down is an invariant transform of a 3D object, it's a rotation and/or translation, an object transformed that way (the system of simplified planes) its the same in visuo-spatial voxel terms. (Also, if you ask a little child what's this and show her a strange chair that's upside-down, she may not tell you a "chair"). And since you understand classical physics (everybody does in one extent or another) and those transforms, you can see how to transform the chair (if laid on the ground) so that you may sit on it as usual. For average chairs/stools, there are some frequent additional features, such as 4 legs, or legs, or a pillar with wheels, or a lever for adjusting the height etc. also there is context - where this "thing" is - in front of a desk, a table, ... (another plane) where people are seen to sit. Also recognition is possible, because parts of chairs are different enough from parts of other objects (otherwise it'd be ambibuous). What means "smashed" - to what extent? If there are other parts which match, they are still recognized, the plane is still there (if the legs still stand, or if the chair is in pieces), the legs are still legs, the back is still back (a plane/surface), the sizes are still the same and fit an average or someone's bottom, there are wheels on a pillar with appropriate size etc >Your method doesn’t add up to a coherent form – a “uniform block” - that all these chairs have in >common – or a coherent method of sitting on them, and identifying them that way. Todor: Mere text explanations are not a formal working system, the details are gotten from the data, and it's faster than explaining it detail by detail. Whatever, there's one way to prove and show what I'm talking about, it's not by fruitless discussions and explanations. -- * Todor "Tosh" Arnaudov * -- Twenkid Research: http://research.twenkid.com -- Self-Improving General Intelligence Conference: http://artificial-mind.blogspot.com/2012/07/news-sigi-2012-1-first-sigi-agi.html -- Todor Arnaudov's Researches Blog: http://artificial-mind.blogspot.com AGI | Archives | Modify Your Subscription ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
