How to guarantee that refined depths have the exact same value at those valid pixels in sparse depth map?

[mqchen1993]

raw_depth_input = blur_depth
...
if sparse_depth is not None:
result_depth = (1 - sparse_mask) * result_depth + sparse_mask * raw_depth_input

你好，我有一个疑惑。为了保留sparse depth上的有效值，为什么不是用sparse上的值替换，而是用blur_depth。blur_depth是网络预测的，这个不一定准确啊？请麻烦帮我解答一下，谢谢！

[mqchen1993]

@XinJCheng

[UltronAI]

I have the same confusion. And I have tried to use
result_depth = (1 - sparse_mask) * result_depth + sparse_mask * sparse_depth
which is as your description in your paper (or it's a mistake?), but in this case, the model can learn nothing.

And it seems that the sparse_depth is only used to compute this sparse_mask, and the model will complete the depth map based on the estimated coarse depth (raw_depth_input in your codes). Why does it work well for depth completion?
@XinJCheng

[sHaol0w]

@UltronAI @mqchen1993 I think it's pointed out in the paper sec 3.3.1 clearly with the exact update equation and explanation ?
The sparse_depth is embedded to a hidden space(blur_depth [at sparse_mask position] ).