project_points() usage

[David-Yan1]

Hi! I was trying to use project_points() to get back a depth map from a modified point cloud. I noticed that there was a black grid pattern on the outputted depth map - not sure if I did something wrong or if it's intended. Below is a depth prediction (predictions["depth"]) for an image, and then the project_points() depth map using predictions["points"].

[lpiccinelli-eth]

Are you resizing the pointmap? That black grid is due to missing points falling in those regions and it could be due to some nearest interpolation

[David-Yan1]

Ah, I believe I was using the incorrect intrinsics in the original. However, the depth is still noisy. Here is a minimal script to reproduce

import open3d as o3d
from PIL import Image
import numpy as np
import torch
from unidepth.utils import colorize
from unidepth.models import UniDepthV1
from unidepth.utils.visualization import save_file_ply
from unidepth.utils.geometric import project_points

model = UniDepthV1.from_pretrained("lpiccinelli/unidepth-v1-vitl14")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)

rgb = np.array(Image.open('assets/demo/rgb.png'))
rgb_torch = torch.from_numpy(rgb).permute(2, 0, 1)
predictions = model.infer(rgb_torch)
depth1 = predictions["depth"].squeeze().cpu().numpy()

depth_pred_col = colorize(depth1, vmin=0.01, vmax=10.0, cmap="magma_r")
Image.fromarray(depth_pred_col).save("original_depth.png")

H, W = rgb.shape[:2]
torch_intrinsic1 = predictions["intrinsics"] 
pcd =  predictions["points"].view(1, 3, -1)  # Shape: (B, 3, H,W) -> (B, 3, HW)
pcd = pcd.transpose(1, 2)  # Shape: (B, HW, 3)

depth = project_points(pcd, torch_intrinsic1, (H,W))
depth_pred_col = colorize(depth.squeeze().cpu().numpy(), vmin=0.01, vmax=10.0, cmap="magma_r")
Image.fromarray(depth_pred_col).save("projected_depth.png")

[lpiccinelli-eth]

Thank you for diving deeper! We have never tried this sanity check :sweat_smile: . There are different possible explanations: 1) The project points may have some problem: some pixels are re-projected on other pixels, thus creating that pepper noise, it may be something related to rounding. 2) The generate_rays is slightly shifting points (hence like rounding effect) 3) spherical_zbuffer_to_euclidean presents some numerical errors, we will try with float64. 4) The interpolation used for depth creates that inconsistency. However, the quasi-random nature of that noise makes me propend for the first/second option.

We will investigate better and try to solve, or at least explain, the source of the problem.

[David-Yan1]

Figured it out! I was rewriting my own projection script and the noise appeared when i used round() instead of using int(). So I believe

# To pixels (rounding!!!), no int as it breaks gradient
    points_2d = points_2d.round()

should be

 # To pixels (rounding!!!), no int as it breaks gradient
 points_2d = points_2d.int()

This fixes the noise as seen below (the comment seems to be wrong).

[lpiccinelli-eth]

Thank you for your comment, your suggestion has been included in the PR #38 which includes V2 release, too