[ICLR 2025 🔥] TempMe: Video Temporal Token Merging for Efficient Text-Video Retrieval

MSRVTT

For MSRVTT, the official data and video links can be found in link.

For the convenience, the splits and captions can be found in sharing from CLIP4Clip,

wget https://github.com/ArrowLuo/CLIP4Clip/releases/download/v0.0/msrvtt_data.zip

Besides, the raw videos can be found in sharing from Frozen in Time, i.e.,

wget https://www.robots.ox.ac.uk/~maxbain/frozen-in-time/data/MSRVTT.zip

Train on MSR-VTT

We conduct experiments on 4 A100x40G GPUs on MSR-VTT, in 2.2.0+cu118 Pytorch.

bash scripts/MSRVTT.sh