I will share my guide with you shortly. But I believe first step would be using the onnx version or tensorRT. Speeds up inference upto 5 times

Skipping many of the frames(using a very low frame rate) and using the A100 GPU helped speed up my feature extraction process.

For me - skipping frames, and extracting in features parallel (limited by GPU memory) - have to make sure that each process has its own instance of the model.

You can use something similar to below to speed up model.track() significantly, you'll loose some tracking results for faster moving cars because of frame skipping, but I believe it can work very well in feature engineering pipelines.

cap = cv2.VideoCapture(video_path)
original_fps = cap.get(cv2.CAP_PROP_FPS)
print(original_fps)
cap.release()

# 1 FPS
vid_stride = 20

# 2 FPS
vid_stride = 10

desired_fps = 4

vid_stride = original_fps / desired_fps
print(vid_stride)

results = model.track(
    source=video_path,
    save=True,
    save_txt=True,
    verbose=False,
    batch=32,
    vid_stride=int(vid_stride)   # <-- 1 FPS (use 10 for 2 FPS)
)

Skipping many of the frames(using a very low frame rate) and using the A100 GPU helped speed up my feature extraction process. google

https://www.google.com

--- HTML ---

google' rel='noopener noreferrer nofollow' target='_blank'>https://www.google.com">google

--- BBCODE ---

[URL=https://www.google.com]google[/URL]

[url=https://www.google.com]google[/url]

[url=https://www.google.com]google[/url]

--- RARE ---

[link=https://www.google.com]google[/link]

[link name=google]https://www.google.com[/link]

((https://www.google.com))

((https://www.google.com))

[https://www.google.com]

[[https://www.google.com]]

[L=google]https://www.google.com[/L]

"google":https://www.google.com