220 likes | 361 Views
Computer Vision Scene Classification Using Neural Nets and a Knowledge Base. Daniel Vevang. Object Detection. Object Detection. Object Detection Training. Object Detection. Object Detection Training. Positive Samples. Object Detection. Object Detection Training. Positive Samples.
E N D
Computer Vision Scene Classification Using Neural Nets and a Knowledge Base Daniel Vevang
Object Detection Object Detection Training
Object Detection Object Detection Training Positive Samples
Object Detection Object Detection Training Positive Samples Negative Samples
Object Detection Object Detection Training Positive Samples Negative Samples Vector Data
Object Detection Object Detection Training Positive Samples Negative Samples Vector Data XML Haarcascade tree
Object Detection Object Detection Training Positive Samples Negative Samples Vector Data XML Haarcascade tree OpenCV Output: Object location and scale from an image.
Object Detection Data: location and scale Scene Detection
Object Detection Data: location and scale Kohonen Network Scene Detection Scene Detection
Object Detection Data: location and scale NN Training Input and Output Data Kohonen Network Scene Detection Scene Detection
Object Detection Data: location and scale NN Training Input and Output Data Trained Kohonen Net Kohonen Network Scene Detection Scene Detection
Object Detection Data: location and scale NN Training Input and Output Data Trained Kohonen Net Kohonen Network Scene Detection Knowledge Base Scene Detection
Tools: OpenCV • Diverse set of computer vision tools
Objectmarker • GUI for Creating a text file of bounding box coordinates for a database of images • Additional scripting tools for creating haar xml cascades. • Eyepatch: Advanced scripting tool for training object detectors. • Warning! Stability Issues!
Kohonen Net Implementation • Code modified from Karsten Kutsa • Still in the process of creating the data model for Neural Net input. • Currently looking to create 8 input nodes for each image (8*5 images) for 40 images total.
Kohenen Net Implementationfor detected images A-E Example input
Parameters to work with • Learning rate for Kohonen layer • Learning rate for output layer • Learning rate for step sizes • Smoothing factor for score deltas • Parameter for width of neighborhood
Additional data to consider • x y location • scale of each object • Multiples of the same object
Knowledge base • Possible implementation of Narl to augment the performance of the Neural Net.