- Ideal clusterizer is done. Center-of-Gravity is calculated on-the-fly for digits which belong to the same cluster. Maxima and Center-of-Gravity are calculated for all digits with the same MC label and in a (time, pad)-window of (16,6) to find single maxima for looper tracks
- Post-processing for the neural network: Remove points where (maximum_pad-CoG_pad + maximum_time-CoG_time) > 2. This is done not to confuse the network (outlier-removal) and it anyway captures most of the distribution of points.
- Efficiency (total maxima found by clusterizer / total maxima found by ideal clusterizer), fake-rate (how many maxima were found by clusterizer which are not found in the ideal clusterizer) and clone-rate (how many maxima are found by ideal clusterizer for which more than one maximum is found by real clusterizer) of maxima is important and ongoing.
- Maximum finding in 3D might improve results: Needs some thought on how to pass values to network (pad offsets and crossing sectors / ROCs)