A video dataset has adjacent frames from the same recording. How should train-test leakage be reduced?

1. Copy all recordings into both partitions
2. Split by recording or independent source unit
3. Randomly split every neighboring frame independently
4. Split only by the order of model filenames