Gaming for Boundary: Elastic Localization for Frame-Supervised Video Moment Retrieval (SIGIR 2025)

Official PyTorch implementation of GOAL, a frame-supervised VMR framework for elastic boundary localization via a game-based paradigm and Dynamic Updating Technique (DUT).

Authors

Hao Liu¹, Yupeng Hu¹*, Kun Wang¹, Yinwei Wei¹, Liqiang Nie²

¹ School of Software, Shandong University, Jinan, China
² School of Computer Science and Technology, Harbin Institute of Technology (Shenzhen), Shenzhen, China * Corresponding author

Paper and Links

• Paper (SIGIR 2025): https://doi.org/10.1145/3726302.3729984

Table of Contents

• Updates
• Introduction
• Highlights
• Method Overview
• Repository Structure
• Installation
• Dataset Preparation
• Checkpoints
• Training
• Evaluation
• Results
• Citation
• Acknowledgement
• License
• Contact

Updates

• [04/2026] Initial public code release.

---

Introduction

This repository contains the implementation of GOAL: Gaming fOr elAstic Localization for Frame-Supervised Video Moment Retrieval (SIGIR 2025).

Frame-supervised Video Moment Retrieval (VMR) aims to retrieve the temporal moment in a video that matches a natural language query using a single annotated frame. GOAL addresses ambiguous boundary retrieval with a game-based paradigm (KFP/AFP/BP) and Dynamic Updating Technique (DUT), improving boundary decisions through unilateral and bilateral updates.

This repository currently provides:

• training code for frame-supervised VMR
• evaluation code for benchmark datasets
• checkpoint files for quick reproduction
• utilities for dataset preparation and experiment management

---

Highlights

• Supports three standard datasets: activitynetcaptions, charadessta, tacos.
• Includes training and evaluation scripts.
• Includes checkpoint folders for quick reproduction:
  • Cha_ckpt/ (I3D / VGG)
  • TACoS_ckpt/

---

Method Overview

---

Repository Structure

|- assets/
|  |- pipeline.png
|- src/
|  |- config.yaml
|  |- dataset/
|  |  |- dataset.py
|  |  |- generate_glance.py
|  |  |- generate_duration_glance.py
|  |- experiment/
|  |  |- train.py
|  |  |- eval.py
|  |- model/
|  |  |- model.py
|  |  |- building_blocks.py
|  |- utils/
|  |  |- utils.py
|  |  |- vl_utils.py
|- README.md

---

Installation

1. Clone the repository

git clone https://github.com/iLearn-Lab/SIGIR25-GOAL.git
cd GOAL

2. Create environment

python -m venv .venv
source .venv/bin/activate   # Linux / Mac
# .venv\Scripts\activate    # Windows

3. Install dependencies

pip install numpy scipy pyyaml tqdm

---

Dataset Preparation

1. Prepare features and raw annotations

This repository follows ViGA's dataset preparation protocol for:

• ActivityNet Captions
• Charades-STA
• TACoS

2. Check path configuration

This repo currently contains local paths (for example under data/...), mainly in:

• src/config.yaml
• src/utils/utils.py

Before running, replace them with your local dataset root and feature paths.

---

Checkpoints

The cloud links of checkpoints: Google Drive and Hugging Face.

---

Training

ActivityNet Captions:

python -m src.experiment.train --task activitynetcaptions

Charades-STA:

python -m src.experiment.train --task charadessta

TACoS:

python -m src.experiment.train --task tacos

---

Evaluation

Evaluate a trained experiment folder:

python -m src.experiment.eval --exp path/to/your/experiment_folder

The folder should contain:

• config.yaml
• model_best.pt

---

Results

Charades-STA (I3D)

Method	Venue	R@0.3	R@0.5	R@0.7	mIoU
ViGA	SIGIR 2022	71.21	45.05	20.27	44.57
CFMR	MM 2023	-	48.14	22.58	-
SG-SCI	MM 2024	70.30	52.07	27.23	46.77
GOAL (Ours)	SIGIR 2025	71.42	55.56	27.69	47.70

Charades-STA (VGG)

Method	Venue	R@0.3	R@0.5	R@0.7	mIoU
ViGA	SIGIR 2022	60.22	36.72	17.20	38.62
D3G	ICCV 2023	-	41.64	19.60	-
GOAL (Ours)	SIGIR 2025	61.16	42.69	21.53	40.56

Charades-STA (C3D)

Method	Venue	R@0.3	R@0.5	R@0.7	mIoU
ViGA	SIGIR 2022	56.85	35.11	15.11	36.35
GOAL (Ours)	SIGIR 2025	56.96	37.74	17.02	36.74

ActivityNet Captions (C3D)

Method	Venue	R@0.3	R@0.5	R@0.7	mIoU
ViGA	SIGIR 2022	59.61	35.79	16.96	40.12
CFMR	MM 2023	-	36.97	17.33	-
D3G	ICCV 2023	58.25	36.68	18.54	-
GOAL (Ours)	SIGIR 2025	59.19	38.72	18.57	40.22

---

Citation

@inproceedings{liu2025gaming,
  title={Gaming for Boundary: Elastic Localization for Frame-Supervised Video Moment Retrieval},
  author={Liu, Hao and Hu, Yupeng and Wang, Kun and Wei, Yinwei and Nie, Liqiang},
  booktitle={Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year={2025},
  doi={10.1145/3726302.3729984}
}

---

Acknowledgement

• The dataset preparation and organization in this repository follow ViGA.
• Thanks to all collaborators and contributors of this project.

---

License

This project is released under the Apache License 2.0.

---

Contact

If you have any questions, feel free to contact me at liuh90210@gmail.com.