Realtime Object Recognition Using Decision Tree Learning Implemented with a Sony AIBO Robot CS 510 Presentation Chris Jorgensen
Realtime Object Recognition Using Decision Tree Learning Implemented with a Sony AIBO Robot CS 510 Presentation Chris Jorgensen
Presentation 1.Abstract/Introduction 2.Problem setup 3.Use of decision tree learning 4.Results 5.Summary/Thoughts
Abstract/Introduction Object recognition –Machine learning used to overcome issues: Domain-specific Complexity inestimable Quality of results –Steps Digital image scanned for features Combine features into “meaningful” attributes Attribute classification
Introduction Continued… Object Recognition Flow
Preprocessing “Obvious” features –Colors –Limbs/Head Shapes derived from image –Used for feature extraction
Problem Setup Recognition –Iterate through surfaces Head, Side, Leg –Generate segments for each surface –Store segments in memory 180 degree memory takes into account camera angle
180 Degree Memory
Problem Setup Continued Segmentation only done on “relevant” pixels –Determined by color Attribute generation* –Color, # segments, # corners, et al –Continuous values discretization via brute- force generated optimal split
Use of Decision Tree Learning Classification via Decision Tree Learning! –Algorithm creates a tree consisting of the attributes; leafs are “symbols” head, side, leg, body, et al –Tree is built by calculating attribute with the highest entropy (depends on # occurrences of each value) –Over-fitting solved by X 2 -pruning Determine if attribute really detects a pattern
Results
Results Continued
Decision Tree Learning –Classification (27 ms) “quite fast” –84% precision on 1080 examples for 5 classes –Even a low number of examples (25) resulted in over 50% precision –Room for improvement noted
Summary/Thoughts Short/vague paper Why do they need faster than 27 ms recognition time? Aibos are slow! Other work on Aibos done at PSU NWCIL –Lendaris/Holmstrom –Aibo uses limb angles, model of motion, to change gait based on floor surface –GA used to generate ideal gait for each surface