I don’t know if it’s a bug or feature in the SDK, but when I call zed.grab() to retrieve frames from an SVO file, it skips some frames if there is a slow process within the while loop. In other words, it doesn’t wait for the process. A good test is to apply a 1-sec sleep and then count the number of frames.
That part of my code is also something like this:
frame_count = 0
frame_count += 1
err = zed.grab(runtime_params)
image_net = image_left_tmp.get_data()
cudnn.benchmark = True img = img_preprocess(image_net) with torch.no_grad(): pred = model(img)
Even if I move the prediction in a separate thread, this still happens.
I’ve tested it on both Linux and Windows. Tested with SDK versions of 3.7 and 3.8.
Is there any way that I can process all the frames in my video in this case?