Skip to content

Latest commit

 

History

History
54 lines (43 loc) · 2.15 KB

README.md

File metadata and controls

54 lines (43 loc) · 2.15 KB

Temporal Sentence Grounding with Relevance Feedback in Videos

This is implementation for the paper "Temporal Sentence Grounding with Relevance Feedback in Videos" overview

Table of Contents

Environments

  • Ubuntu 20.04
  • CUDA 11.7
  • Python 3.7

Install other required packages by

pip install -r requirements.txt

Datasets

This paper has reconstructed the validation and test sets of two widely used datasets in the TSG domain: Charades-STA and ActivityNet Captions, to construct a testing environment for TSG-RF task., i.e., Charades-STA-RF, ActivityNet Captions-RF. The reconstructed dataset is located in the ./data/dataset directory.

Preparation

The details about how to prepare the Charades-STA, ActivityNet Captions features are followed previous work: VSLNet Datasets Preparation. Alternatively, you can download the prepared visual features from Mega, and place them to the ./data/features/ directory. Download the word embeddings from here and place it to ./data/features/ directory.

Training

Train

# train RaTSG on Charades-STA-RF dataset
bash charades_RF_train.sh
# train RaTSG on ActivityNet Captions-RF dataset
bash activitynet_RF_train.sh

Evaluation

Run the following script to test on the trained models: Test

# test RaTSG on Charades-STA-RF dataset
bash charades_RF_test.sh
# test RaTSG on ActivityNet Captions-RF dataset
bash activitynet_RF_test.sh

We release several pretrained checkpoints, please download and put them into ./ckpt/