Maskable Retentive Network for Video Moment Retrieval

Source code for our ACM MM 2024 paper

Task Example: The goal of both MR tasks NLMR (natural language moment retrieval) and SLMR (spoken language moment retrieval) is to predict the temporal boundaries $(\tau_{start}, \tau_{end})$ of target moment described by a given query $q$ (text or audio modality).

 Two important characteristics:
 1) Temporal association between video clips: The temporal correlation between two video clips that are farther apart is weaker;
 2) Redundant background interference: The background contains a lot of redundant information that can interfere with the recognition of the current event, and this redundancy is even worse in long videos.

---

Approach

The architecture of the Maskable Retentive Network (MRNet). We conduct modality-specific attention modes, that is, we set Unlimited Attention for language-related attention regions to maximize cross-modal mutual guidance, and perform a new Maskable Retention for video branch $\mathcal{A}(v\to v)$ for enhanced video sequence modeling.

---

Download and prepare the datasets

1. Download the datasets (Optional).

• The video feature provided by 2D-TAN

    ActivityNet Captions C3D feature
    TACoS C3D feature
  

• The video I3D feature of Charades-STA dataset from LGI

    wget http://cvlab.postech.ac.kr/research/LGI/charades_data.tar.gz
    tar zxvf charades_data.tar.gz
    mv charades data
    rm charades_data.tar.gz
  

• The Audio Captions: ActivityNet Speech Dataset: download the original audio proposed by VGCL

2. For convenience, the extracted input data features can be downloaded directly from baiduyun, passcode:d4yl

3. Text and audio feature extraction (Optional).

 cd preprocess
 python text_encode.py
 python audio_encode.py

4. Set your own dataset path in the following .py file.

  ret/config/paths_catalog.py

5. Or prepare the files in the following structure (Optional).

  MRNet
  ├── configs
  ├── dataset
  ├── ret
  ├── data
  │   ├── activitynet
  │   │   ├── *text features
  │   │   ├── *audio features
  │   │   └── *video c3d features
  │   ├── charades
  │   │   ├── *text features
  │   │   └── *video i3d features
  │   └── tacos
  │       ├── *text features
  │       └── *video c3d features
  ├── train_net.py
  ├── test_net.py
  └── ···

---

Dependencies

pip install yacs h5py terminaltables tqdm librosa transformers
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch

---

Training

ActivityNet

1. python train_net.py --config-file --config-file checkpoints/best/activity/config.yml

TACoS

1. cd ret/modeling/ret_model, then copy the code in file ret_model_tacos.py to file ret_model.py.
2. python train_net.py --config-file checkpoints/best/tacos/config.yml

Charades

please wait for the update

---

Testing

ActivityNet

1. download the model weight file from Google Drive to the checkpoints/best/activity folder
2. python test_net.py --config-file checkpoints/best/activity/config.yml --ckpt checkpoints/best/activity/pool_model_14.pth

TACoS

1. download the model weight file from Google Drive to the checkpoints/best/tacos folder
2. cd ret/modeling/ret_model, then copy the code in file ret_model_tacos.py to file ret_model.py.
3. python test_net.py --config-file checkpoints/best/tacos/config.yml --ckpt checkpoints/best/tacos/pool_model_110e.pth

Charades

please wait for the update

---

LICENSE

The annotation files and many parts of the implementations are borrowed from MMN. Our codes are under MIT license.