Pertinence of evaluation method/metric

[jerome-labonte-udem]

Hi,
I wanted to use you dataset and evaluation method to compare a similar cropping method I am working on.
I made a baseline by cropping a 1:3 and 3:1 aspect ratio window in the middle of each video and was surprise that the results where better than both auto-flip and your improved method on 3:1 aspect ratio (best and mean at least).
Here are the results:
baseline (center-crop) 1:3 : Worst: 43.052 Best: 49.505 Mean: 45.033
3:1 Worst: 73.059 Best: 77.300 mean: 75.409
The 1:3 results are also very close to your first paper results without the clustering part.
I guess the interesting parts are heavily concentrated in the middle of the frames.
What do you think that says about the metric/method ?
Do you know of any other dataset/method that would demonstrate the usefulness of your method or any other ?