jackwayneright's profile - activity

When I check my video frame count using CAP_PROP_FRAME_COUNT, it says there are 300 frames. Similarly, when I use FFMPEG to output the frames of the video, I also get 300 frames. However, when I use a regular VideoCapture read loop similar to the code below, I always get 296 frames. I'm guessing this has something to do with the video format being H264 and the way the data is compressed in such a video. But some of the data I'm using is given with the 300 frame count and my code relies on the OpenCV output which is giving me 296 frames. This discrepancy makes matching the two sets of data a headache. Can anyone explain why there is the difference in frame count/output and how to properly get the two sets of data to match up? In particular, is there just some specific setting I need to look into to get the read() loop to hand back the same 300 frames that FFMPEG would output? Thanks.

Code:

ret, frame = self.read()
frame_count = self.video_capture.get(cv2.CAP_PROP_FRAME_COUNT)
while True:
    if ret:
        # Do stuff with the frame here.
    else:
        break
    ret, frame = self.read()

@StevenPuttemans: Sorry for the delay since the last update. See the latest update above for relevant information.

@StevenPuttemans: Tried it, but with no luck. The Haar training runs fine, so I doubt it's a path issue, though I could be wrong.

@StevenPuttemans: Thanks, but unfortunately, bumping it up to 3GB on each of the buffers doesn't help.

My cascade training works when using Haar but not LBP. The problem seems to occur during the precalcuation phase. For example, when running:

opencv_traincascade -data classifier -vec positive_samples.vec -featureType LBP -bg negative_image_list.txt -precalcValBufSize 1024 -precalcIdxBufSize 1024 -numPos 315 -numNeg 458 -nstages 20 -w 40 -h 40

The output I receive is:

PARAMETERS:
cascadeDirName: classifier
vecFileName: positive_samples.vec
bgFileName: negative_image_list.txt
numPos: 315
numNeg: 458
numStages: 20
precalcValBufSize[Mb] : 1024
precalcIdxBufSize[Mb] : 1024
stageType: BOOST
featureType: LBP
sampleWidth: 40
sampleHeight: 40
boostType: GAB
minHitRate: 0.995
maxFalseAlarmRate: 0.5
weightTrimRate: 0.95
maxDepth: 1
maxWeakCount: 100

===== TRAINING 0-stage =====
<BEGIN
POS count : consumed   315 : 315
NEG count : acceptanceRatio    458 : 1

And it stalls at this point without moving forward (even when waiting for 30+ minutes). If I run this same command using HAAR instead of LBP, the recalculation finishes within 10 seconds or so. I've tried fiddling with the minHitRate and whatnot, but with no other results. When other people's opencv_traincascade stalls, it seems to occur before NEG count : acceptanceRatio is displayed, which leads me to believe I'm having a different problem. Can anyone explain why I might be hitting this wall?

UPDATE

I've found that the program is definitely trying in some way, as it shows up in my Activity Monitor as consuming a huge amount of CPU.

Another user seems to have had this problem a year ago on StackOverflow and tried many things, but they seem to have had no success. Their plight can be found here. It may be worth noting that we are both on OS X.

UPDATE 2

Trying the exact same dataset and command on an Ubuntu machine has the script run correctly. So it seems to be related to the OS X installation of OpenCV in some way.

@berak: That fixes the problem for me! Thanks for the help! I'll look into logging an issue some time in the near future.

Using OpenCV 3.0.0 and Python 3.4.3, I'm getting an error running detectMultiScale3. Specifically the error is:

cascadedetect.cpp:1597: error: (-215) a->size() == n in function clipObjects

Which appears to simply be an assertion check in the C++ source for clipObjects, though I don't know why it would be failing. I'm testing it with the code found at the "Face Detection using Haar Cascades" OpenCV tutorial, but simply changing which version of the detectMultiScale is being used. Anyone know why I might be getting this error and how I can fix it? Thank you much.

detectMultiScale3 Documentation

Code:

import numpy as np
import cv2

face_cascade = cv2.CascadeClassifier('haarcascade_frontalface_default.xml')
eye_cascade = cv2.CascadeClassifier('haarcascade_eye.xml')

img = cv2.imread('face.jpg')
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

faces, rejectLevels, levelWeights = face_cascade.detectMultiScale3(gray, 1.3, 5)
for (x,y,w,h) in faces:
    cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2)
    roi_gray = gray[y:y+h, x:x+w]
    roi_color = img[y:y+h, x:x+w]
    eyes = eye_cascade.detectMultiScale(roi_gray)
    for (ex,ey,ew,eh) in eyes:
        cv2.rectangle(roi_color,(ex,ey),(ex+ew,ey+eh),(0,255,0),2)

cv2.imshow('img',img)
cv2.waitKey(0)
cv2.destroyAllWindows()