[CUDA] opencv cannot decode rtsp video stream

[hexianbin1994]

Expected behaviour

opencv-cuda decode rstp video stream success

Actual behaviour

opencv-cuda can decode mp4 video file, but cannot decode rtsp video stream, with the fault is 【Segmentation fault (core dumped)】

Steps to reproduce

I create docker container and compile opencv-cuda:

docker run -it --network host --gpus all -e NVIDIA_DRIVER_CAPABILITIES=compute,utility,video  nvidia/cuda:12.0.0-cudnn8-devel-ubuntu20.04 bash
## compile ffmpeg
## compile opencv-python
git clone https://github.com/opencv/opencv.git
cd opencv
git checkout 4.8.0
git clone https://github.com/opencv/opencv_contrib.git
cd opencv_contrib
git checkout 4.8.0

compile result is :

NVIDIA CUDA is YES and contain NVCUVID

main code

1 - decode mp4 video file

import cv2

s = "test2.mp4"
cap = cv2.cudacodec.createVideoReader(s)
frame_id = 0
while 1:
    _, img_gpu = cap.nextFrame()
    if frame_id == 100:
        img_cpu = img_gpu.download()
        print(img_cpu.shape)
        break

    frame_id += 1

the result is

2 - decode rtsp video stream

import cv2

s = "rtsp://128.8.84.224:8554/rtsp/test"
cap = cv2.cudacodec.VideoReader(s)
frame_id = 0
while 1:
    _, img_gpu = cap.nextFrame()
    if frame_id == 100:
        img_cpu = img_gpu.download()
        print(img_cpu.shape)
        break

    frame_id += 1

the result is error is in 【_, img_gpu = cap.nextFrame()】

[hexianbin1994]

cv2.getBuildInformation() is

[7576457]

Did you solve the problem?

[asmorkalov]

There were a lot of cudacodec related fixes since 4.8.0. I recommend to update OpenCV to 4.11 and try again. Please also check if stream may be player without CUDA with pure VideoCapture. The issue may happen on FFmpeg side (used by cudacodec for stream parsing, demuxing, etc).