AdaCLIP: Towards Pragmatic Multimodal Video Retrieval

This repo is the implementation for the paper "AdaCLIP: Towards Pragmatic Multimodal Video Retrieval".

Incorporating large image-text foundation models such as CLIP has substantially improved the performance of the multimodal video retrieval task. However, how to practically sample the frames from a video and aggregate the frame features into a video representation is still an open research question. In particular, real-world deployment scenarios, such as embodiment within consumer electronics or cloud-based inference pipelines, require two key facets of retrieval (representation building and search) to be computationally light and fast. In this paper, we propose AdaCLIP, a computationand latency-aware system for pragmatic multimodal video retrieval. AdaCLIP consists of a learning-based frame selection module to select informative frames and a query-independent frame aggregation module to obtain strong video representations from the frame features. Specifically, in the frame selection module, we introduce a differentiable Hard-Top-k algorithm to sample a subset of the frames while optimizing the performance of the video retrieval task in an end-to-end manner. Moreover, to be latency-aware, we also propose a query-independent lightweight approach, MLP-Score, to aggregate the frame features into the video representation, which offers up to 142x speedup on GPU and 822x speedup on CPU in similarity search time compared to query-dependent matching methods. Experimental results on several popular video retrieval datasets confirm the effectiveness of AdaCLIP.

Maintainers

• Angela Ye ([email protected])
• Zhiming Hu ([email protected])

Prerequisites

• Linux (Ubuntu 16.04 or later is recommended)
• Python 3.7
• Packages:
  • ffmpeg ($sudo apt-get install ffmpeg)
• Datasets: ActivityNet Dense Captions, MSRVTT, DiDeMo

How to Install

Create a conda environment and install the appropriate packages:

$ conda create -n adaclip python=3.7 -y
$ conda activate adaclip
$ conda install -y pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.6 -c pytorch -c conda-forge
$ pip install -r requirements.txt

Update cudatoolkit=11.6 with the appropriate CUDA version on your machine.

Datasets

MSRVTT

The videos are shared by Frozen in Time:

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

DiDeMo

The videos can be downloaded from LisaAnne/LocalizingMoments.

ActivityNet

Download the videos from the official website. The authors have made the videos available on Google and Baidu drives.

Preprocessing

Frame Extraction

Run utils/frame_extraction.py after having downloaded the dataset videos and annotations from the website. Make sure that all the videos are in the same directory (no sub-directories allowed).

python utils/frame_extraction.py /path/to/videos /path/to/frames --parallel

Subsequently, update the frames_dir parameter in the config files configs/[dataset].json.

Annotation Preprocessing

If the videos downloaded differ from the set used in the paper, run annot_preprocess/{dataset}_preprocess.py to generate train/test splits used by the dataloader. Splits used in the paper can be found in annots/.

To obtain the annotation files used to generate the splits, please download them from the following links:

• MSRVTT annotations are from CLIP4Clip:

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

• ActivityNet annotations are from the project page of ActivityNet Captions:

wget https://cs.stanford.edu/people/ranjaykrishna/densevid/captions.zip

• DiDeMo annotations have two components: annotations from the original author and the split used by Collaborative Experts.

Training

Train AdaCLIP on MSRVTT

python -m torch.distributed.launch --nproc_per_node=3 train.py --config configs/msrvtt-jsfusion.json --frames_dir /path/to/msrvtt/frames --top_k 12 --freeze_cnn --frame_agg mlp

Train AdaCLIP on DiDeMo

python -m torch.distributed.launch --nproc_per_node=3 train.py --config configs/didemo.json --frames_dir /path/to/didemo/frames --top_k 16 --num_epochs 5 --freeze_cnn --frame_agg mlp

Train AdaCLIP on ActivityNet

python -m torch.distributed.launch --nproc_per_node=3 train.py --config configs/activitynet.json --frames_dir /path/to/activitynet/frames --top_k 32 --freeze_cnn --frame_agg mlp

Train uniform sampling & mean aggregation baseline on MSRVTT

python -m torch.distributed.launch --nproc_per_node=3 train.py --config configs/msrvtt-jsfusion.json --frames_dir /path/to/msrvtt/frames --no_policy

You may need to update the batch size in the config files and/or the number of GPUs used by changing the --nproc_per_node=NUM_GPU parameter.

Inference

Inference on MSRVTT

python -m torch.distributed.launch --nproc_per_node=3 train.py --config configs/msrvtt-jsfusion.json --frames_dir /path/to/msrvtt/frames --top_k 12 --frame_agg mlp --do_inference --resume /path/to/msrvtt/checkpoint

Citation

If you use this code, please cite our paper:

@inproceedings{AdaCLIP,
    title={AdaCLIP: Towards Pragmatic Multimodal Video Retrieval},
    author={Hu, Zhiming and Ye, Angela N and Khorasgani, Salar H and Mohomed, Iqbal},
    booktitle={Proc. of the 31st ACM International Conference on Multimedia},
    year={2023}
}