What is the minimum size for training sets in bagOfFeatures analysis.

I am trying to count two species of geese in 50 megapixel aerial photos. I locate the geese using morphology and the detectSURF function. I then crop the sub-images of individual geese (see examples below) and evaluate the species with a model derived from bagOfFeatures analysis. The probability of correct classification promised to be very good based on the confusion matrices: * The confusion matrix for this test set is:

                                      PREDICTED
                                  | Brant   CAGO   
            KNOWN     Brant      | 1.00     0.00   
                        CAGO      | 0.00     1.00

• Average Accuracy is 1.00 for training set of 30 images for each classEvaluating image category classifier for 2 categories.-------------------------------------------------------
• Category 1: Brants
• Category 2: CAGO
• The confusion matrix for this test set is:

                              PREDICTED
                               | Brant   CAGO   
          KNOWN        Brant  | 0.97     0.03   
                         CAGO    0.01     0.99

• Average Accuracy is 0.98 for validation set of 70 images for each class.

I increased my training set for each species of goose to over 300 for the final model, but when I apply the model to my large images about half of geese that are evaluated are identified for each class. Do I need more training images?