TrajViT

"One Trajectory, One Token": a grounded video-tokenization method from Ranjay Krishna's RAIVN lab (ICCV 2025) that represents video as panoptic sub-object trajectories. TrajViT uses ~10× fewer tokens than space-time patch tokenizers while improving average VideoQA accuracy by 5.2% across six benchmarks and cutting inference FLOPs by 18×.