This seems to be an interesting competition. I am still a bit confused (just the error metric has enough reason to be confused ) I would like to ask some clarification regarding the rules.
"Note: It is strictly prohibited to add additional annotations or hand-label the dataset. Algorithms must be created using the annotations as provided."
Does this mean that annotating additional training frames is prohibited? Obviously the test set should not be in any way annotated. Similar kaggle competition winning solutions often have additional training annotations.
Can we use external data? e.g. ImageNet
Can we use pretrained models? e.g. https://github.com/fchollet/keras/blob/master/keras/applications/vgg16.py