What should I do to make getting objects from video (YouTube) faster?

[KatarinaPopikova]

Hello!
All I need is to get the objects that appeared in the video (YouTube).
So I tried set attribute vid_stride to 50 and imgsz=320, but it takes almost the same time. I also tried device="cpu" and I got the same result in computing time.

I tried this:

model.predict(source=SOURCE, device="cpu", verbose=False)

model.predict(source=SOURCE, device=0, vid_stride=50, verbose=False, imgsz=320)

What should I do to make getting objects from video faster?

[glenn-jocher]

@KatarinaPopikova hello! To speed up object detection in videos from YouTube using YOLOv8, consider the following suggestions:

1. Reduce Frame Rate: Increasing vid_stride skips more frames, which should reduce processing time. Try setting it even higher if you can afford to miss more frames.

2. Lower Resolution: You've already tried reducing imgsz to 320. You might experiment with going even lower, but be aware that this can affect the accuracy of detection.

3. Batch Processing: If YOLOv8 supports batch processing for videos, process multiple frames at once to utilize your GPU more efficiently.

4. Hardware Acceleration: Ensure you're using a GPU if available (device=0 typically refers to the first GPU). The CPU (device="cpu") is generally slower.

5. Optimized Model: Use a smaller and faster model variant if available, which can provide a good balance between speed and accuracy.

6. Parallel Processing: If you're processing multiple videos, you can run them in parallel using multiprocessing techniques to utilize all CPU/GPU resources.

7. Profile Your Model: Use the Benchmark mode to profile your model and identify bottlenecks.

Remember that each video and use case may require a different combination of these optimizations for the best results. 🚀