Closed wyclearnpy closed 3 weeks ago
hi @wyclearnpy, One way of increasing training / pre-processing efficiency is to train on every 2-nd or even 4-th frame of the video. In our updated paper (see Appendix B.2), we show that DINO-Tracker maintains its performance under such compression, while reducing the memory and time requirements 2x or 4x. In other words, DINO-Tracker can generalize to unseen frames from the input video.
What are the specific training commands? how should i set?
When I was conducting experiments with my own data, I ran out of memory during the data preprocessing stage. My data was a ten-second video with thirty frames per second and 1280x1024 pixels. The graphics card is 2080ti and the memory is 22g. How should I adjust it so that the experiment can work properly?