[SOLVED] python opencv cv2.cv.CV_CAP_PROP_FRAME

python opencv cv2.cv.CV_CAP_PROP_FRAME_COUNT get wrong numbers

import os
import cv2
path='/home/nlpr4/video-data/UCF-101/GolfSwing/v_GolfSwing_g24_c06.avi'
cap=cv2.VideoCapture(path)
video_length=int(cap.get(cv2.cv.CV_CAP_PROP_FRAME_COUNT))

success=True
count=0
while success:
    success,image=cap.read()
    if success==False:
        break
    count=count+1

print video_length,count

output:

149 
146

why the two numbers different? what's wrong?

Solution

The get() for CAP_PROP_FRAME_COUNT is never supposed to be accurate! If you check the opencv source code. You can find this:

int64_t CvCapture_FFMPEG::get_total_frames() const
{
    int64_t nbf = ic->streams[video_stream]->nb_frames;

    if (nbf == 0)
    {
        nbf = (int64_t)floor(get_duration_sec() * get_fps() + 0.5);
    }
    return nbf;
}

This means it will first look into the stream header for nb_frames, which you can check with ffprobe. If there is no such field, then there is no better way to get frame number than directly decoding the video. The opencv did a rough estimation by get_duration_sec() * get_fps() + 0.5 which surely not mean for accuracy.

Thus, to obtain the correct frame number you have to decode and read through the entire stream, or you have to ask the video generator to generate correct stream header with nb_frames field.

It is unfortunately to see after almost 10 years opencv still does not correct the implementation: we should use av_read_frame(fmt_ctx, &pkt) to read and parse all packet without actual decode them.

I personally stopped relies on opencv for video decoding more than 10 years ago, but recently I have to collaborate with some project using opencv to do so. I have to call ffprobe in a subprocess to get the correct frame counts:

ffprobe -v error -select_streams v:0 -count_packets -show_entries stream=nb_read_packets -of csv=p=0 input.mp4