MobileViCLIP: An Efficient Video-Text Model for Mobile Devices

Our paper is available in MobileViCLIP

News

[2025.10.8] The code is updated.
[2025.6.28] Our MobileViCLIP has been accepted by ICCV2025. The code will be updated soon.

Environment preparation

1: Create environment

conda env create -f requirements/mobileviclip.yml

2: Activate environment

conda activate mobileviclip

Data preparation

1: Download videos

All datasets can be downloaded from the official website.

Supposing these videos are in the following path:

data
├── anet_1.3_video_val_resize
	├── _1vYKA7mNLI.avi
	└── .....
├── kinetics_400_val_10s_320p
	├── __lt03EF4ao.mp4
	└── .....
└── MSRVTT_Videos
	├── video0.mp4
	└── .....

2: Download annotations

You can download the annotations of each datasets from Google Drive link. Then, put them in the following path:

anno_downstream
└── InternVid
     └── InternVid-10M-flt_1.json
     ├── InternVid-10M-flt_2.json
     └── .....
├── anet_ret_val.json
└── .....

3: Prepare pretrained weights

We adopt pre-trained MobileCLIP from MobileCLIP. You can also download this model from Google Drive link.

We also provide our mobileviclip_tiny.pt and mobileviclip_small.pt from Google Drive link.

Then, put them in the following path:

checkpoints
└── mobileclip_s0.pt
├── mobileclip_s2.pt
├── mobileviclip_tiny.pt
└── mobileviclip_small.pt

Train and test

Please run the following commad for training and inference

For Training,

bash scripts/pretraining/mobileviclip_<tiny or small>/run.sh

For Inference,

bash scripts/evaluation/clip/zero_shot/mobileviclip_<tiny or small>/eval_<dataset>.sh

Credits

We especially thank the contributors of the MobileCLIP and InternVideo for providing helpful code.