Research Dreams

Markus Vincze (and colleagues)

 

1.         A person's "best friend":

 

It can perceive the mental state of a human to support her/him. E.g., give elderly people impact to do to something, to excite them not to become passive. We see it like to good companion or the old wise grandfather/mother.

 

Related Applications:

-           be a companion: cheer up, find someone to talk to (not the machine itself, also another person), help when needed, suggest help when found necessary, help when detected as really needed -      

            learn favourites and assist (pre-selection for shopping clothes,

...)

 

-           clean up, after cooking, after party, after child, keep order at state wanted and appreciated as comfortable 

-           basic daily needs: cook, look at recipe and shop (find ripe fruits and vegetables, edible mushrooms), assist during cooking, detect ready meat on oven, detect hairs in soup, ... 

-           extension to aid handicapped persons, be "eyes of blind people" (a person's best friend with additional capabilities)   Some needed cognitive capabilities (interesting problem): 

-           detect non-object (danger, fear, emotions, ...) 

-           learn how to react and to handle different emotional states 

-           understand relations between humans and objects  

-           understand relations between two persons 

-           reason about environment and its functions,  

-           recognise dangerous situations (screw loose, hose open, pot at edge) and errors/faults 

-           separate good and not good actions (good in the sense of good for the person and the community rules) 

-           learn alternative behaviours dealing with the same situation (there is not only one way to solve a problem), this ability helps to decrease failures,  additionally if a system is assisting a person maybe preferences for one solution are demanded (learn priority of alternative solutions).

 

Envisioned results of open research problems: 

-           recognise objects under all lighting and cluttered settings possible (many objects, classes of objects, learn objects from a few views) 

-           find/recognise glass (handle reflections, maybe with other sensing modes) 

-           search web for objects/data of interest, e.g., red sport car, baroque furniture 

-           form relations between objects, reason about relations and consequences 

-           learn representations, learn to select a good representation, evolve representations (for fast tracking, recognition, gesture interpretation, ...) 

-           learn expectations and strategies: search for object in a house (not only on one table) = look everywhere, even open drawers, stir in heaps of objects 

-           represent states, situations, models (a generic prediction tool using environment models, e.g., bouncing ball, pot falling of shelf, a human action expected)  

-           memory at different time scales 

-           automated learning of concepts given by humans (as part of acting system) 

-           learning of concepts by observation, maybe arriving at concepts different to humans, and in this way find out more about how humans represent concepts 

-           recognise and interpret sign language (or any gestures) and find differences between cultures, automatically translate gesture for another cultural environment 

-           recognise und understand intentions, draw conclusions from them 

-           evolving vision system (learn by observation, keep running all time) 

-           vision system that works and knows it works and knows it, or subparts, fails (=does not fulfil the task), asks another vision system to help in case of failure 

-           serve multiple purposes and resolve conflicts 

-           formalized methods for high level reasoning, linking high level reasoning with low level vision, reason about elements of low level vision 

-           camera/sensor working independent of lighting conditions 

-           cooperating cognitive vision systems 

 

Solving the research issues helps to solve other applications, for example: 

-           the perfect sport referee: detect fouls, incorrect play in all games, offside in football, ... 

-           sport aids: hit baseball, golf (+ golf caddy), play flipper, teach player correct motions, ... 

-           understand hand writing, hand drawings 

-           recognise forged paintings, writings 

-           detect weather conditions, earthquakes, hurricanes, floods

 

2. Community Care Taking System

 

A system to aid society, a system that has a "broad understanding" of humanity, a system that brings people together and makes sure nobody is harmed. A visionary role as "human protector", similar to Giskard in Asimov's novels. (Though mind reading is maybe too far away.) Example applications, starting from simple and leading to more complex scenarios: 

-           detect child seat on front seat to not unlock airbag 

-           intelligent traffic control switching on most demand 

-           detect waste, separate waste 

-           detect landmines 

-           detect unusual behaviour of patients under observation in hospital 

-           detect drunken people getting into the car, car does not start 

-           detect situations before they lead to accidents and react accordingly 

-           find thieves (while they are doing it, better stop them before) 

-           detect and catch insects 

-           detect bullets and move protection to avoid persons being hit

 

Why interesting problem: Takes the "personal friend" scenario a step further and reasons not about individual humans but about groups of persons, their behaviour, the relation to other groups, or the relation within a mass of persons.  It requires the understanding and interpreting the actions and emotions of many to many person relations (also involving objects). Hence, it adds complexity to the needed cognitive abilities above.

 

Envisioned results of open research problems: Same as above with the extension to groups, plus group specific topics: 

-           understand and derive group behaviour from individual behaviour 

-           understand inter-group behaviour 

-           representation of abstracted group/human values and morale 

-           multiple cooperating vision systems 

-           detect animals (dolphin in net and set it free)