This repository aims at generating relative poses from camera transformation matrices, using the help of neural network architecture.
The temporal attention layer added is helpful to acheive higher accuracy. At the moment the network is trained on one of the 7-scenes dataset generating about 60% accuracy.