SEMANTIC PART BASED OBJECT DETECTION

Coert van Gemeren

Cognitive Artificial Intelligence, Utrecht University

Abstract

This page is about an ongoing master-thesis project performed by Coert van Gemeren under supervision of dr. Robby T. Tan and Zhouyu Fu. The goal of this project is to create a part based multi-view multi-object detector that can achieve state of the art performance. The object class models are trained by a semi-unsupervised learning process, which generates a generic multi-view object model which represents an object class. The system we are building can learn object classes from large annotated image sets such as the The PASCAL Visual Object Classes data collection in an unsupervised manner. Employing a latent-SVM algorithm a hierarchical multi-view object model is trained. After training several different class models, an attributed relational graph is generated which correlates the parts of the model to the parts of other models, based on shape similarities.


Temporary Results:

1. Multiple car detections from a still image

The detection result of applying a learned model to an image with multiple cars.

Below are 3 of the 6 root filter models from a mixture model trained on car images from the PASCAL Object Recognition Database Collection 2010. The other 3 models are the mirrored equivalents of the shown models. This approach follows Felzenszwalb, et al's approach very closely, the main difference being that this system is implemented completely in C++ rather than a combination of Matlab and C.

2. Further Work: Part correlation

Currently we are conceiving a mechanism that will correlate similar parts of different object class models to each other. To do this we are investigating the use of attributed relational graph matching.

3. Face detection on a webcam

Below is a small video showing a simple implementation of a part-based star-model for object detection (in this case of faces). This work is based on Fergus et al..


References:

  1. Felzenszwalb, P. F. and Girshick, R. B. and McAllester, D., Object Detection with Discriminatively Trained Part Based Models.
    IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 32, No. 9, September 2010.
  2. Fergus, R. and Perona, P. and Zisserman, A., Weakly Supervised Scale-Invariant Learning of Models for Visual Recognition
    International Journal of Computer Vision, Vol. 71(3), 273-303, March 2007.
  3. The PASCAL Visual Object Classes Homepage, website.
  4. Zhu, L. and Chen, Y. and Torralba, A. and Freeman, W. and Yuille, A., Part and Appearance Sharing: Recursive Compositional Models for Multi-View Multi-Object Detection
    IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). 2010.

Last Update: February 2011