search

fangchangma / self-supervised-depth-completion

ICRA 2019 "Self-supervised Sparse-to-Dense: Self-supervised Depth Completion from LiDAR and Monocular Camera"
MIT License
623 stars 135 forks source link

Why do you need to multiply the result of final conv by 100? #20

Closed AbnerCSZ closed 5 years ago

AbnerCSZ commented 5 years ago

In the end of model.py


        if self.training:
            return 100 * y
        else:
            min_distance = 0.9
            return F.relu(100 * y - min_distance) + min_distance # the minimum range of Velodyne is around 3 feet ~= 0.9m

Could your tell me what is the 100 means?

fangchangma commented 5 years ago

Could your tell me what is the 100 means?

It was legacy code and removing the multiplier should not affect the training much. However, the learning rate might require some adjustment.

AbnerCSZ commented 5 years ago

Thank you, by the way, why not add smooth loss in the dense d mode(args.w2 = 0 when mode is d). Have you compared the difference?

fangchangma commented 5 years ago

Thank you, by the way, why not add smooth loss in the dense d mode(args.w2 = 0 when mode is d). Have you compared the difference?

I believe I did the experiment but did not observe improvement. Please let me know if you observe otherwise.