Incomplete data files in provided dataset — missing right images and disparity maps

[ShihuaHuang95]

Hi, thanks for your great work and for releasing the dataset!

We’ve downloaded the dataset following the official instructions step by step. However, during verification, we noticed that many samples are missing either the right images or disparity. Here’s a quick summary:
.
Our Questions:
Is this expected, or are some files actually missing from the dataset you provided?
Could you kindly confirm whether our download is complete or corrupted?
If possible, could you provide MD5 / SHA256 checksums for the original compressed files (e.g., .tar.gz) so we can verify the integrity of our downloads?

Thanks again for your contributions — and looking forward to your clarification!

[David-Yan1]

Thanks for pointing that out! Your download appears fine, these are just scenes that crashed during datagen but weren't filtered in post-processing. You can safely just remove or ignore these pairs during data-loading.

On a slightly related note, we noticed that there are some ill-posed or very difficult cases in the current dataset due to a bug in the post-processing code (mostly clipped objects). We will re-release the dataset as soon as possible, but you can filter out these pairs yourself by recomputing the depth and thresholding pairs with objects that are too close. Here's an example helper script that you can modify. We personally used a threshold of 0.125m for indoor/flying scenes and 5m for nature scenes, but the numbers (particularly for nature) can be tweaked.

import numpy as np

def camera_center(T: np.ndarray) -> np.ndarray:
    R, t = T[:3, :3], T[:3, 3]
    return -R.T @ t                      

def baseline_vector(T1: np.ndarray, T2: np.ndarray) -> np.ndarray:
    return camera_center(T2) - camera_center(T1)  

def baseline_length(T1: np.ndarray, T2: np.ndarray) -> float:
    return np.linalg.norm(baseline_vector(T1, T2))

def depth_from_disparity(disp_px: np.ndarray, f_px: float, B: float) -> np.ndarray:
    with np.errstate(divide="ignore", invalid="ignore"):
        Z = np.where(disp_px > 0, f_px * B / disp_px, 0.0)
    return Z.astype(np.float32)

def main():
    K_L = np.load('indoor/7c3e6e1c/frames/camview/camera_0/camview_2_0_0048_0.npz')['K']
    K_R = np.load('indoor//7c3e6e1c/frames/camview/camera_1/camview_2_0_0048_1.npz')['K']
    T_L = np.load('indoor/7c3e6e1c/frames/camview/camera_0/camview_2_0_0048_0.npz')['T']
    T_R = np.load('indoor/7c3e6e1c/frames/camview/camera_1/camview_2_0_0048_1.npz')['T']

    B = baseline_length(T_L, T_R)         
    disp = np.load('/n/fs/scratch/dy2617/release/indoor/7c3e6e1c/frames/disparity/camera_0/disparity_2_0_0048_0.npy')

    f_px = K_L[0, 0]             
    depth = depth_from_disparity(disp, f_px, B)

    reject = np.any(depth < 0.125)) # We use 0.125 m for indoor scenes, 2.5-5m for outdoor scenes


if __name__ == "__main__":
    main()

[ShihuaHuang95]

Thanks for confirming this! Please do let me know if there are any updates in the future. Really appreciate it — great work!