Closed LiuJF1226 closed 2 months ago
Thank you for pointing that out! I'll fix it.
Setting bg_color
to [0,0,0,0,0,0,0,0,0,0,0,0] here may be helpful. But this raises another problem. Directly setting the background for features and uncertainty to zero, or binding a learnable background for features and uncertainty, which one could be better?
It may depend on the type of scene.
For bounded and simple scenes, such as synthetic scenes with a dark background, setting the background to all zeros might be helpful since the decoder doesn’t need to fit a complex distribution.
For unbounded and large scenes, such as urban environments where the background has specific meaning, having a learnable background for both features and uncertainty might be more effective in fitting the decoder’s distribution.
It is just my intuition, if there is any mistake or problem, please point it out and discuss it more.
Thank you for your reply. For generalization ability, I think it is more reasonable to employ a learnable background for both features and uncertainty. But It needs more experiments for validation.
Thank you for the interesting work. I notice that
NUM_CHANNELS
is modified to 12. But thebg_color
is still [0,0,0] of 3 channels. Therefore, this line can lead to array index out of bounds forbg_color
.