search

tinghuiz / SfMLearner

An unsupervised learning framework for depth and ego-motion estimation from monocular videos
MIT License
1.96k stars 555 forks source link

multiplication by a small constant #101

Closed aviziskind closed 5 years ago

aviziskind commented 5 years ago

In the nets.py, in the calculation of pose, you multiply by a small constant of 0.01:

                # Empirically we found that scaling by a small constant 
                # facilitates training.
                pose_final = 0.01 * tf.reshape(pose_avg, [-1, num_source, 6])

I am curious how you came to discover this... and if you have any intuition as to why this should help to facilitate training?

tinghuiz commented 5 years ago

The intuition was that since we are sampling consecutive frames, the pose change should be relatively small. So multiplying by a small constant incorporates such prior information.