Very bad quality compared to Metric3D

[seamie6]

I am running this on nuscenes with the following code adapted from your demo.py file:

import numpy as np
import torch
from PIL import Image

from unidepth.models import UniDepthV1, UniDepthV2
from unidepth.utils import colorize, image_grid

from nuscenes.nuscenes import NuScenes
import os


def demo(model):
    rgb = np.array(Image.open(image_path))
    rgb_torch = torch.from_numpy(rgb).permute(2, 0, 1)

    # predict
    predictions = model.infer(rgb_torch, intrin)

    # get GT and pred
    depth_pred = predictions["depth"].squeeze().cpu().numpy().astype(np.uint8)
    
    depth_pred_PIL = Image.fromarray(depth_pred)
    depth_pred_PIL.save('test.png')

    depth_pred_col = colorize(depth_pred, vmin=0.01, vmax=100.0, cmap="magma_r")
    im = Image.fromarray(depth_pred_col)
    
    im.show()



if __name__ == "__main__":
    dataroot = '../nuScenes/'
    nusc = NuScenes(version='v1.0-mini', dataroot=dataroot, verbose=True)

    samples = nusc.sample
    cams = ['CAM_FRONT_LEFT', 'CAM_FRONT', 'CAM_FRONT_RIGHT', 
            'CAM_BACK_RIGHT', 'CAM_BACK', 'CAM_BACK_LEFT']

    rec = samples[0]
    cam  = 'CAM_BACK'
    samp = nusc.get('sample_data', rec['data'][cam])
    imgname = os.path.join(nusc.dataroot, samp['filename'])

    sens = nusc.get('calibrated_sensor', samp['calibrated_sensor_token'])
    intrin = torch.Tensor(sens['camera_intrinsic'])

    image_path = imgname
    
    name = "unidepth-v2-vitl14"
    # model = UniDepthV1.from_pretrained("lpiccinelli/unidepth-v1-vitl14")
    model = UniDepthV2.from_pretrained(f"lpiccinelli/{name}")

    # set resolution level (only V2)
    # model.resolution_level = 0

    # set interpolation mode (only V2)
    # model.interpolation_mode = "bilinear"

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = model.to(device)
    demo(model)

which just runs the model on one back camera image. Here is the results:

METRIC3D GIANT MODEL:

UNIDEPTH LARGE MODEL:

From your paper, it claims to outperform Metric3D but here it is performing quite bad, the result is extremely blurry and quite innacurate So I assume I am doing something wrong and could you point me in the right direction? Thanks

[lpiccinelli-eth]

Thank you for raising this concern. The snippet you showed looks fine, although I do not know about the intrinsics format: you can try not providing any camera to be on the safe side.

However, here are a few points to consider:

1. If you want to reproduce the paper results, you should use Metric3Dv1 and UniDepthv1, those were the public models at writing time.
2. Usually we compare models with the same backbone. In this case, Metric3D uses giant while UniDepth uses large. Moreover, Metric3D uses a larger image shape as input and we did not have any "sharpness" loss hence better details.
3. In (metric) depth estimation, fine-grained details and global correctness are seldom positively correlated, i.e. visually appealing depth maps are not necessarily geometrically correct. Therefore, you should provide quantitative results on multiple scenes before drawing any conclusions on accuracy.

I would be happy to answer any further concerns and additional findings of yours.

[seamie6]

Thank you for the quick reply. I implemented Metric3D into a model and it performed quite well, I will try to do the same for UniDepth and see the results, and then draw my conclusions. Perhaps when viewed as an image it can be deceiving, thank you.