About the number of input video frames

[elvindp]

Hi, when l use the example python api to inference a video with 15mins, it consume large GPU memory and report torch.cuda.OutOfMemoryError. I find the script "load_video" extract 1 frame per second, which need 900 frames for 15min video and cost large memory. So is there any suggested number of input frames for long videos? for example, one video 30 frames?

[CarlCloudWang]

sample some frames per second... def load_video(video_file, skip=1): from decord import VideoReader vr = VideoReader(video_file)

# Get video frame rate
fps = vr.get_avg_fps()

# Calculate frame indices for 1fps
frame_indices = [int(fps * i) for i in range(int(len(vr) / fps))]
frames = vr.get_batch(frame_indices).asnumpy()
ls = []

if skip == 1 or skip is None:
    ls = [Image.fromarray(frames[i]) for i in range(int(len(vr) / fps))]
elif skip > 1:
    ls = [Image.fromarray(frames[i]) for i in range(int(len(vr) / fps))][::skip]  # sample frame
elif skip < 0:
    ls = [Image.fromarray(frames[i]) for i in range(int(len(vr) / fps))][:-skip]  # skip<0 and get 0 to -skip frame

return ls