Master-Praktikum - Sensorgestützte intelligente Umgebungen (IN2106, IN8902, IN4053)
Dauer | 6 SWS |
Art | Praktikum |
Semester | WS2011 |
Vortragende | Zoltan Marton, Nico Blodow |
Lehrziel | siehe Modulbeschreibung IN2106 |
Sprache | Englisch |
Termin | 22.11.2011, 3PM |
Midterm presentations | January 13, 2011 |
Endterm presentations | March 16, 2011 |
Project Ideas
Interactive Segmentation (Dejan Pangercic)- segmentation of everyday objects in the cluttered scenes
- inclusion of further cues such as color, 3D features, manipulator feedback
- multi-cue classification
- Research Areas: image segmentation, control theory, machine learning.
- Tools: Linux, ROS, OpenCV, PCL
CUDA-based 3D Image Processing for Object Recognition (Nico Blodow)- multiple-cue image segmentation and classification
- combination of image and depth features to increase performance / robustness
- Conditional Random Field (or similar) classifier
- Research Areas: computer vision, machine learning, 3D processing
- Tools: Linux, ROS, OpenCV, PCL (pcl::cuda)
- Ensemble Methods for Multi-Cue Object Classification (Zoltan-Csaba Marton)
- combination of vision- and 3D-based features from the Kinect sensor
- ensemble of specialized routines to improve object detection
- building an object model database for storing information about objects
- perception system for selecting the right methods to use for a given task
- Research Areas: machine learning, cognitive vision, semantic mapping
- Tools: Linux, ROS, your brain (and possibly PCL and OpenCV)
- Eye-in-Hand Modeling (Dejan Pangercic)
- enabling the technology for eye-in-hand-based reconstruction of scenes
- design of the multi-purpose tool holder
- hand-eye calibration for the external tool using AR markers
- simple scene reconstruction from multiple views
- Research Areas: Image recognition, control theory, mapping
- Tools: Linux, ROS, OpenCV, PCL, CNC Machine
- Hand-held Modeling of Indoor Scenes (Dejan Pangercic)
- RGBD SLAM (simultaneous localization and mapping using RGB and depth cameras)
- evaluate existing RGBD SLAM approaches (UW, Freiburg, CCNY, Willow Garage)
- depending on previous findings improve or combine with IAS tools
- scene reconstruction of various (actual) households
- IKEA example with interior design assistance
- Research Areas: SLAM, mapping, surface reconstruction, computer graphics
- Tools: Linux, ROS, OpenCV, PCL.
- Semantic Labeling of Places (Dejan Pangercic)
- build an interface to Mechanical Turk for annotation of objects
- build interface for creation of ontologies of objects
- propose abstraction layer classify scenes using various ML classifiers (WUP, Binary Trees, SVM, CRF)
- Research Areas: webGL, machine learning, knowledge representation
- Tools: Linux, ROS
Barcode-based Object Recognition (Dejan Pangercic)- build a barcode-based object recognition system (in collaboration with Barcoo)
- get grasp of the object detection and pick-up system
- find/enable web camera ROS driver with autofocus support
- barcode finding algorithm
- create an object ontology
- Research Areas: drivers, image recognition, control theory, knowledge representation
- Tools: Linux, ROS, OpenCV, Android
Memory modeling for object perception (Nico Blodow)- higher-level approach to remembering object detection & locations
- key problem: remember where objects are, what they look like
- loosely modeled after human memory (short term / long term memory)
- per-location map integration of memory, per-object collection of visual sensory inputs
- Research Areas: cognitive robotics, artificial intelligence
- Tools: Linux, ROS
- Detecting small objects in cluttered scenes (Nico Blodow)
- build a system that can detect a small set of objects in possibly cluttered scenes using 2D and 3D image processing
- combination of depth / surface cues and RGB cues (e.g. color, edges)
- in this project, objects will be BRIO child toys (wooden train track pieces)
- Research areas: object detection
- Tools: Linux, ROS, OpenCV, PCL (possibly CUDA)
Multi-view CAD-model matching to Kinect data (Zoltan-Csaba Marton)- start off with am RGB-Depth frame sequence from the Kinect camera
- detect objects (and their approximate pose) – using an existing method
- find the exact pose of the object by model fitting (use existing method)
- extend system to account for the information about empty regions of space
- extension to keep/improve multiple hypotheses over time (see particle filter)
- Tools: Linux, ROS, PCL (possibly OpenCV)
- 3D object tracking system (after detection) (Nico Blodow)
- Problem: When grasping, robot might move the object before establishing a stable grasp
- 2D tracking methods often give poor 3D poses
- combination of conventional tracking with depth reprojection + occlusion handling
- Research areas: object tracking, computer vision
- Tools: Linux, ROS, OpenCV, PCL
- Context-aware Object Recognition Architecture (Zoltan-Csaba Marton)
- no general architecture for knowledge-enabled perception system
- several sub-parts are implemented already in open-source projects
- related architectures are implemented in related field (e.g.Wattson - the question answering program that won Jeopardy)
- survey existing architectures (UIAM, AI-Goggles, Proctor, MOPED,BLORT) and highlight their advantages
- based on the lessons learned, create a prototype for an object recognition system that makes decisions about an object's types given heterogeneous information
- Research Areas: knowledge representation, cognitive vision, identity resolution
- Tools: Linux, ROS, partially Java, and a bit of creativity :P
- Real-Time Point Cloud Annotation in RViz (David Gossow)
- Build a system that allows a user to select 3D regions of interest in a point cloud
- Use SLAM to accurately track the position of the selected region while the robot is moving
- Collect training data (Point clouds/labeled depth+RGB images) from that
- Train object recognition algorithms on training set
- Evaluate algorithms on annotated point cloud test set
- Research areas: computer vision, 3D processing, user interfaces
- Tools: Linux, PCL, OpenCV, ROS, RViz+Interactive Markers, C++
A Combined Geometry & Texture Based Interest Point Detector (David Gossow)- Extend an existing algorithm to jointly detect salient texture and geometry features (blobs)
- Combine with (existing) feature descriptor, e.g. BRIEF
- Evaluate for tasks such as object recognition, RGBD-SLAM
- Research areas: computer vision, 3D processing
- Tools: Linux, OpenCV, PCL, C++
- Fast & Reliable Face Detection using a Kinect Camera (David Gossow)
- Extend the current standard face detection algorithm (Viola/Jones) in OpenCV to exploit depth information
- Use existing intensity-based classifier from OpenCV
- Write a detector that, instead of constructing a scale pyramid, directly infers the classification window size at each image position from camera parameters and depth information
- Make use of integral images to efficiently compute mean filters of variable scale
- Create/use existing annotated test set & evaluate classification performance & speed against image-only approach
- Research areas: computer vision, face detection, machine learning
- Tools: Linux, OpenCV, C++
Detecting and Picking-up Cups and Plates in the Dishwasher (Thomas Rühr)- Research areas: computer vision, 3D processing, PR2
- Tools: Linux, PCL, ROS, C++