When we did the annotation process for the Zackenberg flower pictures I insisted in annotating the WHOLE flower. In my mind, the more information we gave the algorithm, the better it would perform. This means that we have some annotations that contain flowers in strange positions. For example we have annotation with flowers that are in a half-moon position. In this case the flower is contained in the corners of the annotation and the middle is “background”. Actually, in some annotation, given the position of the flower, most of the information is background.
Reading Navneet Dalal’s paper and thesis on Histogram of Oriented Gradients I realized that on of the assumptions of the whole process is that the element that is being detected – pedestrians in Navneet Dalal’s case and flowers in mine should be more – should occupy more or less the same place in all annotations. In the case of pedestrian detection, all were in an upright position.
All this leads me to believe that we should be careful when doing annotations for the flowers. We should strive to “normalize” the way the flower appears in the annotation. So, instead of annotating the “half-moon” flower once, we can annotate it twice trying to keep each sub-annotation in the centre. For now, this is just a hunch. I’ll test with both types of annotation and see what the results are.