ModelsWeights's picture
Upload folder using huggingface_hub
ebb66db verified

Boolean masks of the UBnormal Human-Related subset

The folders testing and validating contain the boolean masks for the test set and the validation set, respectively.

Within each folder, a hr_mask.npy is provided, which contains the masks of the clips concatenated by lexicographic order, i.e. in the same order as their names are listed in the folder.

The folders {testing, validating}/test_frame_mask contain the masks for each clip as separated Numpy files. Values within the masks are boolean such that

mask[i] = True if the frame should be kept
mask[i] = False otherwise

For each set, a stats.json file shows the total number of frames in the set, the number of discarded frames and the percentage of discarded frames; the same information is also reported for each clip.

To use the HR mask, add in the evaluation code the following:

import numpy as np

HR_UBNORMAL_MASK = np.load('data/UBnormal/hr_bool_masks/testing/hr_mask.npy')

if the clips in the evaluation are loaded and appear in the code in the same order as their names are sorted in the folder, otherwise

import numpy as np

# example for the clip 001_0100
HR_UBNORMAL_MASK_001_0100 = np.load('data/UBnormal/hr_bool_masks/testing/test_frame_mask/001_0100.npy')

for each clip.