neurips26evalmde/scenes

This is the dataset we created for studying behaviors of different metrics, as mentioned in Sec 3 of our paper. Each folder in it corresponds to 1 scene. Each scene has the following structure

SCENE_DIR/
 - gt_depth.parquet
 - rgb.png
 - valid_triangle.parquet
 - synth_depth/
 --- affine_transformation_of_depth/
 ----- ...
 --- boundary/
 ----- ...
 --- curv_high_freq/
 ----- ...
 --- relative_scale/
 ----- ...
 --- affine_transformation_of_disparity/
 ----- ...
 --- camera_intrinsics/
 ----- ...
 --- curv_low_freq/
 ----- ...
 --- surface_orientation/
 ----- ...

Where,

• gt_depth.parquet, rgb.png are ground truth depth and rgb image of the scene.
• valid_triangle.parquet is a reference file we provide for generating Textureless Relighting visualization. Please refer to README.md of our code for the meaning of this file.
• Each subfolder under synth_depth contains depth perturbed by different intensities of one perturbation type. The mapping between subfolder name and perturbation type can be found below. The file name of each perturbed depth stands for the perbutation intensity, which can be found in Supp. A of our paper, except for curvature perturbations, where the float before _ in the file name stands for perturbation intensity and the float after _ in the file name stands for $\sigma$.

subfolder name	affine_transformation_of_depth	boundary	curv_high_freq	relative_scale	affine_transformation_of_disparity	camera_intrinsics	curv_low_freq	surface_orientation
perturbation type	affine transform of depth	boundary	curvature (high frequency)	relative scale	affine transform of disparity	camera intrinsics	curvature (low frequency)	surface orientation

Please refer to README.md of our code for a detailed description of depth file format.

Our code loads depth and valid_triangle from .npz files. But due to better compatibility of .parquet to hugging face, we converted our .npz files to .parquet when uploading them. To use our code on this dataset, we provide load_parquet.py which loads .parquet files in our dataset as a dictionary of numpy arrays. The loaded dictionary is compatible with our code, i.e. can directly replace np.load($NPZ_FILE) in our code. Install python environment for load_parquet.py by pip install numpy==2.0.0 pyarrow==24.0.0.