Researchers from Information Technologies Institute (CERTH-ITI) have deployed support vector regression as a technique to model the performance of a visual recognition model based on the quality and the quantity of its training samples. More specifically, starting from a classifier which was trained on an initial set of images, the regressor can predict whether the addition of user tagged images can boost the performance of the initial classifier. The regressor takes into account both the quality of the new images that is hindered by the noise in their associated tags and the maturity of the existing model which can be saturated if the original training set is of large size already. The demonstrated experimental results look promising, as shown in the figure. Here, the performance prediction regressor is compared to various baselines. We can see that the proposed regressor (magenta line) performs much closer to the optimal upper baseline (red line) than using a random predictor.
This is particularly important in the content of the Live+Gov requirement for mobile visual recognition, as the end users can now more effectively and efficiently train their own visual recognition models, since they have the additional knowledge of how many images they need. This novel scientific work will be presented in one of the most prestigious conferences on image processing in the following October .
 E. Chatzilari, S. Nikolopoulos, Y. Kompatsiaris, J. Kittler. “How many more images do we need? Performance Prediction of bootstrapping for Image Classification.”, in the 21st IEEE International Conference on Image Processing (ICIP 2014), Paris, France, October 27-30, 2014.