Padding inputs with zeros before prediction causes affinity predictions to wrong

Mohinta2892 / catena

Neuron Segmentation, Synaptic Partner Detection and Microtubule tracking for vEM with EM-2-EM translation. Codebase built upon Funke lab's algorithms.

7 stars 1 forks source link

Padding inputs with zeros before prediction causes affinity predictions to wrong #6

Open Mohinta2892 opened 7 months ago

Mohinta2892 commented 7 months ago

Input raw EM that are smaller than (232,232,232) in shape need to be padded before inference. However, if padded with 0s, predicted affinities may not be highly accurate (see screenshot).

Potential solution is to use reflect padding, instead of constant in np.pad