InterpoNet, A brain inspired neural network for optical flow dense interpolation Supplementary

We found that the training process results declined when the average number of missing pixels in the training flow maps was too high. Some of the matching algorithms, in particular DeepMatching, did produce sparse maps like these. To tackle this problem, we calculate the flow map bi-directionally (From I to I ′ and from I ′ to I) using the matching algorithm. We invert the second flow map and average the two maps. This simple step solves the sparseness problem for all of the matching algorithms we used. This procedure added to the computation time of our method. However most matching algorithms already compute bidirectional maps for consistency check and false matches filtering purposes and so we did not need to apply them twice. Importantly, we found that the bi-directional averaging is critical mostly for training the network and specifically for DeepMatching outputs. Training on FlowFields non averaged maps, for instance, gives comparable results to training with the averaged maps. Interestingly, applying EpicFlow on the bi-directional average of the DeepMatching algorithm output also slightly improved their results (Table S.1). For consistency reasons, we choose to present in this paper the results gained using the bi-directional averaged maps for training and evaluation. However, for all matching algorithms using only the original, non averaged map, in evaluation time yields results similar to those presented (Table S.1). The analysis in this section was performed without the variational post processing for both our method and EpicFlow.