thanks @LukasBommes and his project mv-extractor for the idea

There is already CAP_PROP_LRF_HAS_KEY_FRAME property for the I frames.
Do we really need detailed B/P frames too?

I mean in order to use CAP_PROP_LRF_HAS_KEY_FRAME you need to read the video in raw mode. Also, in some tests the output of CAP_PROP_LRF_HAS_KEY_FRAME actually did not result in correct I-frames information for me with a mpegts file.

I mean in order to use CAP_PROP_LRF_HAS_KEY_FRAME you need to read the video in raw mode. Also, in some tests the output of CAP_PROP_LRF_HAS_KEY_FRAME actually did not result in correct I-frames information for me with a mpegts file.

I also meet this problem.
I receive some rtsp streams by raw mode , sometims the CAP_PROP_LRF_HAS_KEY_FRAME never return true. In order to decode more rtsp streams, I want to only decode I-frames. So how detect I-frame more reliably under raw mode?

Does the property CAP_PROP_FRAME_TYPE not work for you? Did you try something like this:

cap = cv2.VideoCapture(...)
ret, frame = cap.read()
if cap.get(cv2.CAP_PROP_FRAME_TYPE) == ord("I"):
    # do something with this I-frame

Yes, it work. However the read function decode every frame. I just wan to decode I-frame.
I use cv::cudacodec::VideoReader and implement interfacecv::cudacodec::RawVideoSource to decode, so I need use raw mode VideoCatpure to receive raw data and send only I frame data to VideoReader