GantMan
Detect nsfw in video file
Hello. Thank you very much for this excellent work. I'm using it to detect nsfw in video frames. The problem I have is that I can only do it if I extract a frame, save it to disk and then open it and analyze it. I use opencv. Is there any way to directly analyze the frame that gives opencv ?
Let me give you an example
cam = cv2.VideoCapture('video.avi') while(True): ret,frame = cam.read() result = detector.predict(frame) ???????
We have a ticket to do it in JavaScript that might be applied soon. But I'm actually not that great at Python. I imagine if you could grab all the keyframes down to a folder, then run all those images in the folder in a single batch, that would be pretty fast.
@GantMan I think it'd be inaccurate by simply grabbing all keyframes, since keyframes can be defined by producer during transcoding. Therefore, it's possible that we get mislead by producer. On the other hand, given we have so many frames, in case of false positive, we will have a hard time to decide whether it's bad or not. I think we need a model dedicated for video handling
What I do is take an image every n frames, see if it is nsfw and if positive I keep it in a directory
I keep them as well, and then I look at the directory to check what was analyzed. In the tests I have done, I have found quite a few false positives.
This is the code to decide if is or not nsfw: ` result = detector.predict('temp.jpg') datos = result['temp.jpg']
nsfw = False
if float(datos['porn']) > float(datos['drawings']) and float(datos['porn']) > float(datos['hentai']) and float(datos['porn']) > float(datos['neutral']) and float(datos['porn']) > float(datos['sexy']) and float(datos['porn']) > 0.93:
nsfw = True
else:
if float(datos['sexy']) > float(datos['drawings']) and float(datos['sexy']) > float(datos['hentai']) and float(datos['sexy']) > float(datos['neutral']) and float(datos['sexy']) > float(datos['porn']) and float(datos['sexy']) > 0.93:
nsfw = True
if nsfw:
name = './' 'nude_' + info_video[0] + '/' + info_video[0] + '_' + str(currentframe) + '.jpg'
print ('Creating...' + name + ' --- ' + str(float(datos['porn'])) + ';' + str(float(datos['sexy']))+ ';' + str(float(datos['neutral'])))
cv2.imwrite(name, frame)
frames_nude += 1
currentframe += 1`
@xcbtrader based on your condition, you will have false negatives. Moreover, you will treat hentai as good content for work.
@xcbtrader Yeah, saving frames to disk and then calling the predict function is very time taking. I will ad a function to process a video directly. I will also add a paramter in that function so that you can select how many frames you want to analyze. eg: n=1 will analyze all frames and n=0.5 will analyze only half of randomly selected frames each second. Would this work for you?
I do it differently. I put a variable called n_frames that indicates how many frames it analyzes. For example: If n_frames = 10, it reads frame by frame, and every 10 frames it analyzes. Analyze frame 10, 20, 30, 40 ....
@xcbtrader I think you need to define a model to decide whether the video is nsfw based on discrete frame detection output. Moreover, your fixed frames can be cracked if someone intends.
@xcbtrader Yeah, saving frames to disk and then calling the predict function is very time taking. I will ad a function to process a video directly. I will also add a paramter in that function so that you can select how many frames you want to analyze. eg: n=1 will analyze all frames and n=0.5 will analyze only half of randomly selected frames each second. Would this work for you?
Has this been implemented yet? I haven't found any reference to this function.
Right now I'm extracting frames from a video with ffmpeg, and sending them in batch to the predict method. I've found that doing this with up to 5 videos at once really kills the CPU. I'm using a 4 core CPU with 8 gigs of RAM. Any recommendation on the hardware requirements part for such a task for analyzing multiple videos at once
Nothing in place yet. I'd love to see your solution. This is something we might come back to as more advancements show up in TFJS