This is Tensorflow implementation of DispNet from https://lmb.informatik.uni-freiburg.de/Publications/2016/MIFDB16/MIFDB16.pdf
Currently DispNet and DispNetCorr1D are implemented and converge in experiments.
Model | FlyingThings3D EPE | FPS (K40) |
---|---|---|
DispNetCorr1D (slow) | 2.28 | 5 |
DispNetCorr1D (fast) | 3.4 | 10 |
There are two versions of custom Correlation1D layer: one is built upon simple tf operations like pad and slice in "for" loop - that is the slow one. Faster implementation utilizes the CUDA code from original Caffe layer and allows to inference the model in exactly the same time as in Caffe, but this implementation is not fully tested yet.