why the frame is not equal with opencv?

[buaacarzp]

why the frame is not equal with opencv? my test code as below,please give me an answer,thank you!

from decord import VideoReader,cpu,gpu,VideoLoader 
import cv2 
import numpy as np
def frame_reader(video_path):
    video_capture = cv2.VideoCapture()
    if not video_capture.open(video_path):
        video_capture.release()
        return
    ret = 1
    cnt = 0
    res = [] 
    while ret:
        if cnt >=5:
            break
        ret,frame = video_capture.read()
        res.append(frame)
        cnt +=1
    video_capture.release()
    return res
res = frame_reader('/data/zhoup/workspace/videoDecode/examples/preview.mp4')
vr = VideoReader('/data/zhoup/workspace/videoDecode/examples/preview.mp4', ctx=cpu(0))
lists = [0,1,2,3,4]
frames = vr.get_batch(lists).asnumpy()
for i in range(len(lists)):
    v1 = res[i]
    v2 = frames[i]
    equal = (v1==v2).all()
    print(equal)

[jbohnslav]

what's the codec on the video file you're reading? I've found that decord does not match imageio for MJPG videos-- seems the JPG decoder is not identical

[seaman1900]

I met the same problem, most dataset use the OpenCV to load pcitures and use them to train their model, however, I got the different array with Decord and OpenCV, what's more, the model got the wrong result (human pose estimation)

[saandeepa93]

Was this issue resolved? I am getting different frame length on decord vs OpenCV and Torchvision.

[NeighborhoodCoding]

You are using wrong conversion code. It should be converted dmlc_img from decord.ndarray.NDArray to numpy.ndarray. Also, OpenCV is generally using BGR, not RGB. You should chage this part also!

Try this recipe.

cv2_img = cv2.cvtColor(cv2_img, cv2.COLOR_BGR2RGB)
print((cv2_img == dmlc_img.asnumpy()).all())

Hope this helps. Also GPU HW acceleration is also very experimental. It is highly recommended that you turn off this feature.

True