This is the official repository of the paper "Video Moment Retrieval from Text Queries via Single Frame Annotation" published in SIGIR 2022.
https://arxiv.org/abs/2204.09409
This project has been tested on the following conda environment.
$ conda create --name viga python=3.7
$ source activate viga
(viga)$ conda install pytorch=1.10.0 cudatoolkit=11.3.1
(viga)$ pip install numpy scipy pyyaml tqdm
This repository contains our glance annotations already. To replicate our work, one should prepare extra data and finally get the following structure.
ckpt/ our pre-trained model, available at https://drive.google.com/file/d/1S4e8XmIpiVFJKSSJ4Tig4qN0yaCwiVLs/view?usp=sharing
data/
+-- activitynetcaptions/
| +-- c3d/
| +-- annotations/
| | +-- glance/
| | | +-- train.json
| | | +-- val_1.json
| | | +-- val_2.json
| | +-- train.json downloaded
| | +-- val_1.json downloaded
| | +-- val_2.json downloaded
+-- charadessta/
| +-- i3d/
| +-- c3d/
| +-- vgg/
| +-- annotations/
| | +-- glance/
| | | +-- charades_sta_train.txt
| | | +-- charades_sta_test.txt
| | +-- charades_sta_train.txt downloaded
| | +-- charades_sta_test.txt downloaded
| | +-- Charades_v1_train.csv downloaded
| | +-- Charades_v1_test.csv downloaded
+-- tacos/
| +-- c3d/
| +-- annotations/
| | +-- glance/
| | | +-- train.json
| | | +-- test.json
| | | +-- val.json
| | +-- train.json downloaded
| | +-- test.json downloaded
| | +-- val.json downloaded
glove.840B.300d.txt downloaded from https://nlp.stanford.edu/data/glove.840B.300d.zip
Downloaded from http://activity-net.org/challenges/2016/download.html. We extracted the features from sub_activitynet_v1-3.c3d.hdf5 as individual files.
Folder contains 19994 vid.npys, each of shape (T, 500).
Downloaded from https://cs.stanford.edu/people/ranjaykrishna/densevid/
We extracted the C3D features of Charades-STA by ourselves. We decided not to make it available for downloading due to our limited resource of online storage, and the fact that this process can be easily replicated. To specify, we directly adopted the C3D model weights pre-trained on Sports1M. The extracted feature of each clip in a video was the fc6 layer, and the clips were sampled via sliding window of step size of 8, window size of 16. Our codes for this extraction were based on this repo. https://github.com/DavideA/c3d-pytorch
Folder contains 9848 vid.npys, each of shape (T, 4096).
Downloaded from https://github.com/JonghwanMun/LGI4temporalgrounding. This is the features extracted from I3D (finetuned on Charades). We processed them by trimming off unnecessary dimensions.
Folder contains 9848 vid.npys, each of shape (T, 1024).
Downloaded from https://github.com/microsoft/2D-TAN. We processed the data by converting the downloaded version vgg_rgb_features.hdf5 into numpy arrays.
Folder contains 6672 vid.npys, each of shape (T, 4096).
Downloaded from https://github.com/jiyanggao/TALL
Downloaded from https://github.com/microsoft/2D-TAN. We extracted the features from tall_c3d_features.hdf5 as individual files.
Folder contains 127 vid.npys, each of shape (T, 4096).
Downloaded from https://github.com/microsoft/2D-TAN
Our models were trained using the following commands.
(viga)$ CUDA_VISIBLE_DEVICES=0 python -m src.experiment.train --task activitynetcaptions
(viga)$ CUDA_VISIBLE_DEVICES=0 python -m src.experiment.train --task charadessta
(viga)$ CUDA_VISIBLE_DEVICES=0 python -m src.experiment.train --task tacos
Our trained models were evaluated using the following commands.
(viga)$ CUDA_VISIBLE_DEVICES=0 python -m src.experiment.eval --exp ckpt/activitynetcaptions
(viga)$ CUDA_VISIBLE_DEVICES=0 python -m src.experiment.eval --exp ckpt/charadessta_c3d
(viga)$ CUDA_VISIBLE_DEVICES=0 python -m src.experiment.eval --exp ckpt/charadessta_i3d
(viga)$ CUDA_VISIBLE_DEVICES=0 python -m src.experiment.eval --exp ckpt/charadessta_vgg
(viga)$ CUDA_VISIBLE_DEVICES=0 python -m src.experiment.eval --exp ckpt/tacos