Research
160d agoTrajTok: End‑To‑End Video Tokenizer
Apple’s TrajTok replaces patch‑based tokenization with a learnable trajectory tokenizer that adapts to a video’s semantic complexity. By co‑training the tokenizer end‑to‑end with the model, TrajTok eliminates slow external segmentation pipelines and reduces token count, boosting efficiency and scalability for video understanding tasks.
The Signal
This approach also simplifies deployment and improves generalization across diverse video domains.