About realtime inference

[As957LoveYourSelf]

That's a great work!
Now I want to use it for real-time processing of camera captures, but when I tried inference with two frames, the output feature points for both frames showed zero difference. After testing, I found that at least 10 frames of data need to be processed simultaneously. Do you have any good suggestions?

[zbw001]

Hi! Our model is trained using absolute positional encodings with a fixed window length of 16. It may not generalize to smaller window sizes. At the moment, we don't have an optimal method for processing one delta frame at a time. One possible workaround could be to repeatedly run the model using the most recent 16 frames. However, depending on your frame rate, this approach might be too slow for realtime applications.

[As957LoveYourSelf]

Oh, I see. Thank you very much for your response. This is still an excellent work. Wish you smooth progress in your research!