We know SIFT algorithm ( Scale-invariant feature transform) can be used in image classification problem. After getting the SIFT descriptor, we usually use k means clustering to combine multiple descriptors of an image into one feature. This requires the information of all images in the training set to construct the cluster space.

I wonder that if in practice we have learned the classification model (such as SVM), and we want to classify a new image but we don't have the training data (only the model), how can we deal with the descriptors of the new image? K means or PCA is not available in this case.