The Depth Prediction on Waymo Dataset seems not good

[xiao10ma]

I use the unidepth-v2-vitl14 to predict the waymo dataset depth. And it seems that it doesn't predict a good result.
I just change the image_path from "assets/demo/rgb.png" to the waymo path.
The demo you provide is correct, however, the waymo dataset is not.

Here's my code.

from unidepth.models import UniDepthV1, UniDepthV2
import torch
from unidepth.utils import colorize, image_grid
import numpy as np
from PIL import Image

model = UniDepthV2.from_pretrained("lpiccinelli/unidepth-v2-vitl14") # or "lpiccinelli/unidepth-v1-cnvnxtl" for the ConvNext backbone

# Move to CUDA, if any
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)

# Load the RGB image and the normalization will be taken care of by the model
# image_path = "assets/demo/rgb.png"
image_path = '/HDD_DISK/users/mazipei/BZG/dataset/017/images/000000_0.png'
rgb = np.array(Image.open(image_path))
rgb_torch = torch.from_numpy(rgb).permute(2, 0, 1)


predictions = model.infer(rgb_torch)
depth_pred = predictions["depth"].squeeze().cpu().numpy()

depth_pred_col = colorize(depth_pred, vmin=0.01, vmax=10.0, cmap="magma_r")
artifact = image_grid([rgb, depth_pred_col], 1, 2)
Image.fromarray(artifact).save("./res.png")

Here's the result:

Do you have any idea where is the problem? Thank you so much.