Inferior results for DA3-SMALL compared to DepthAnythingV2-small

Hi,

Thank you for publishing the code and models!

I was comparing the performance of smallest variants for DAv2 and DA3 on indoor scenes from DIODE and NYUv2 datasets. I observe incorrect depth predictions (e.g., shadowed regions) with blurry/bleeding edges by DA3-SMALL comapred to DAv2-small, which is seen in benchmarking numbers as well.

Maybe I am missing something here and eager to know your thoughts on it.

Example from DIODE [RGB -vs- DA3-small -vs- DAv2-small]

Example from NYUv2 [RGB -vs- DA3-small -vs- DAv2-small]

The colormaps are inverted between DA3 and DAv2 since they are trained to predict scale-and-shift-invariant depth and inverse depth, respectively. One may need to zoom into the depth maps to see the finer details.

-- Goutam