Unable to reproduce NYUv2 fine-tuning results (97.2% vs reported 98.8% δ1)

[Cats-eat-peppercorns]

Hi,

Thank you for releasing UniDepthV2 — the model works great for zero-shot depth estimation.

I'm trying to reproduce the NYUv2 fine-tuning results from Table V of the paper (arXiv:2502.20110), where UniDepthV2-Large achieves δ1 = 98.8% after fine-tuning on the NYU training set.

What I've done:

Used the published checkpoint pytorch_model_l.bin from this repo as the pretrained model
Evaluated with the UniDepth eval crop [45:-9, 41:-39] and GT camera intrinsics
Verified zero-shot δ1 ≈ 95% (consistent with the paper's SUN-RGBD/IBims-1 results)
My best fine-tuning result: δ1 = 97.2% (1.6% below reported 98.8%)

I've experimented with:

frozen_stages ∈ {0, 6, 12, 18, 24}
lr ∈ {5e-6, 1e-5, 2e-5, 5e-5}
n_iters ∈ {2000, 3000, 5000, 10000, 20000}
Different random seeds
fp32 evaluation (no improvement over fp16)
My question:
Could you share the specific fine-tuning hyperparameters used for the NYUv2 results in Table V? Specifically:

Which checkpoint was used as the starting point — the same pytorch_model_l.bin published here, or a different training-stage checkpoint?
What were the learning rate, number of iterations, and frozen_stages?
Was val_no_gt_camera set to true or false during evaluation?
Thank you for your time!