Dark-48: a dark video dataset for action recognition in the dark.

Dark-48 is a dark video dataset created for action recognition in the dark, collected from Kinetics700 and MiT, contains 8815 dark videos belong to 48 action classes. To overcome the limitations (fewer scenes and fewer categories) of the existing dark video action recognition dataset (e.g. ARID) , we build this dataset to provide more dark videos, more categories and richer scenes.

The full dataset can be downloaded from:

• baidu, code: dk18
• Google Drive

The action categories for Dark-48 data set are:

''' adult+female+singing, adult+male+singing, adult+male+speaking, applauding, ascending, balancing, bouncing, breathing fire, bubbling, burning, cheering, child+singing, clapping, combusting, coughing, dancing, descending, driving, dropping, drumming, erupting, floating, headbanging, juggling, juggling fire, karaoke, performing, playing, playing laser tag, playing+music, playing+videogames, raining, riding mechanical bull, rising, rocking, shooting off fireworks, shouting, silent disco, singing, smoking, spelunking, spilling, spinning, spinning poi, storming, talking, turning, whistling '''

Here the examples for each category:

The source code of the proposed dark video evaluation method for dark videos collection.

import os
import cv2
import numpy as np
from decord import VideoReader

def dark_img(img, threshold = 0.877):
    YCrCb = cv2.cvtColor(img, cv2.COLOR_RGB2YCrCb)
    Y = YCrCb[:,:,0]
    # Determine whether image is bright or dimmed
    exp_in = 112 # Expected global average intensity 
    M,N = img.shape[:2]
    mean_in = np.sum(Y/(M*N)) 
    t = (mean_in - exp_in)/ exp_in
    
    # Check image
    if t < -threshold: # Dimmed Image
        return True
    else:
        return False

def dark_video(video, segments=8, threshold = 0.877):
    vr = VideoReader(video)
    seg = int(len(vr) / segments)
    sample_id = [seg*i+int(seg/2) for i in range(0, segments)]
    frames = vr.get_batch(sample_id).asnumpy()
    video_t = 0
    for i in range(len(sample_id)):
        img = frames[i]
        img_t = dark_img(img)
        video_t += img_t
    if video_t/segments >= threshold:
        return True
    else:
        return False

If you have any questions about the Dark-48 dataset, please contact:

yzliu.me at gmail.com