Unable to get same metric results in zero-shot relative depth estimation

Hi, thanks you all for your great work and open sourcing.

I am working on testing Depth Anything v2 large and Depth Anything v1 large on zero-shot relative depth estimation with DIODE dataset. For Depth Anything v1, I got 0.214 for AbsRel and 0.756 for δ1. For Depth Anything v2, I got 0.218 for AbsRel and 0.755 for δ1. However, in the papers you reported 0.066 and 0.952 for Depth Anything v1 and 0.065 and 0.954 for Depth Anything v2.

I use the following code to align Depth Anything output with the ground truth depth map of DIODE dataset:

inverse_truth = 1.0 / (truth_masked + epsilon)
A = np.concatenate([output_masked, np.ones_like(output_masked)], axis=-1)
scale, shift = np.linalg.lstsq(A, inverse_truth, rcond=None)[0]
alignmented_output = 1.0 / (output_map * scale + shift + epsilon)

As there is such a huge difference in metrics, could you kindly provide your evaluation code on zero-shot relative depth estimation on DIODE and possibly other dataset? Thank you very much!