A question about module 'DN2'.

[freshjh]

Hello, thanks for your nice work. And I have a little question about the role of 'DN2' and the distillation strategy in this paper.
Can we fuse the simulated 3D feature with the 2D feature directly, if the '3D' feature contains more geometric or spatial information? In this way, the simulated 3D feature can be seen as an auxiliary or guidance of 2D feature.
Besides, have you experiment the distillation without DN1 (e.g., constraint the difference between 3D feature and 2D feature with L1 loss)?
Looking foreword to your reply. Thanks!^_^

[liuzhengzhe]

Thanks for your message.

1. Of course we can use other feature merging approaches, such as summation, concatenation..
2. Without DN1, the 2D feature can also be trained to be similar to 3D features. DN_1 is to facilitate this process.