Semi-supervised multiple instance learning based domain adaptation for object detection

• Authors: Chhaya Methani;Rahul Thota;Amit Kale
• Venue: Proceedings of the Eighth Indian Conference on Computer Vision, Graphics and Image Processing
• Year: 2012

In this paper we propose a semi-supervised multiple instance learning based boosting algorithm for domain adaptation, with face detection as an example. Very often a generic classifier learned using a large volume of training data needs to be tuned to work for a specific scenario. However when deployed, the test scenarios may differ marginally from the training ones. For e.g. a face detection system may be deployed in an airport as well as in an auditorium hallway. The classifier then needs to adapt to the new domain. Instead of retraining the classifier completely using examples from the new scenario, it is desirable to see how much the classifier can "self-learn". Conventional self-learning algorithms consider putative positives on test data given by the base classifier, and select a subset of those based on more stringent thresholds. In this paper we propose an alternative self-learning approach which is based on the popular multiple instance learning approach which makes use of "bags" instead of single instances for training the classifier. We pool the putative positives on a given test image into a positive bag and the putative negatives into a negative bag. We augment this data to the initial training data and retrain the classifier using MILBoost. Specifically the advantage of our approach is that since it makes use of bags it is more robust to classification errors by the base classifier. We demonstrate the improvement in classification accuracy using our approach on Faces in the Wild database. We show that our approach outperforms self-learning and compares favorably with MILBoost trained on manually marked face data without the corresponding increase in labeling effort.