Official PyTorch implementation of “Generation-Augmented Video Corpus Moment Retrieval”, published at the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2026).
- Authors: Mingjin Kuai, Qianyin Xiao, Juncheng Li, Jin Peng, Lizi Liao, and Wei Ji
- Paper: ACM Digital Library
- Code: github.com/QianyinXiao/VideoGAR
Video-GAR addresses video corpus moment retrieval (VCMR) with a Bi-Mamba backbone and a training-only generative query-reconstruction objective. The generative decoder regularizes video-text alignment during training and adds no decoder overhead at inference time.
The released inference path uses the model's learned start/end boundary predictors directly. An experimental post-processing scorer that was not used in the paper is not part of this release and is not required to train or evaluate Video-GAR.
The implementation is based on ReLoCLNet. Dataset features and evaluation data follow the setup described in the paper and the ReLoCLNet repository.
Video-GAR requires Python 3.9–3.11 and a CUDA-capable GPU. The environment follows the dependency baseline used by the authors' related VERSE repository.
| Package | Version |
|---|---|
| PyTorch | 2.4.0 (CUDA 11.8) |
| torchvision | 0.19.0 |
| mamba-ssm | 2.3.0 |
| transformers | >=4.38.2, <5 |
| tokenizers | >=0.15.2, <0.20 |
| NumPy | >=1.24, <2 |
| h5py | >=3.10, <4 |
| TensorBoard | >=2.15, <3 |
| easydict | >=1.13, <2 |
| tqdm | >=4.66, <5 |
The complete resolved environment is recorded in uv.lock. On Linux x86-64, the lock uses the official prebuilt mamba-ssm wheel for Python 3.9–3.11, so a local CUDA toolkit and nvcc are not required for installation.
The recommended installation uses uv and the CUDA 11.8 wheels declared in pyproject.toml.
git clone https://github.com/QianyinXiao/VideoGAR.git
cd VideoGAR
pip install uv
uv python install 3.9
uv sync --extra gpu-cu118 --extra bimamba
source .venv/bin/activate
source setup.shIf PyTorch is already managed by your cluster, install the core dependencies with uv sync and install a PyTorch/mamba-ssm build compatible with the cluster's CUDA toolchain separately.
Verify the installation and run the synthetic GPU smoke tests with:
python -m unittest discover -s tests -vThe GPU tests are skipped automatically when CUDA or the optional Bi-Mamba dependencies are unavailable.
VideoGAR/
├── method_core/
│ ├── model.py # VideoGARModel and fusion encoder
│ ├── bimamba.py # Bidirectional Mamba encoder layer
│ ├── query_decoder.py # Training-only query decoder
│ ├── config.py # Training and evaluation options
│ ├── train.py # Training entry point
│ └── inference.py # VCMR/SVMR/VR inference
├── evaluation/
│ └── eval.py # Retrieval metrics
└── utils/ # Data, tensor, and model utilities
- Bi-Mamba backbone: efficient bidirectional sequence modeling for long video contexts.
- Generative augmentation: query reconstruction acts as an auxiliary training objective.
- Fusion encoder: provides query-conditioned video memory to the training-only decoder.
- Boundary-aware moment localization: predicts start and end boundaries for retrieved moments.
If this repository or the paper is useful in your research, please cite:
@inproceedings{kuai2026generation,
author = {Mingjin Kuai and Qianyin Xiao and Juncheng Li and Jin Peng and Lizi Liao and Wei Ji},
title = {Generation-Augmented Video Corpus Moment Retrieval},
booktitle = {Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval},
publisher = {ACM},
year = {2026},
pages = {857--868},
doi = {10.1145/3805712.3809742},
url = {https://doi.org/10.1145/3805712.3809742}
}This project is released under the MIT License.
Keywords: Video-GAR, VideoGAR, video corpus moment retrieval, VCMR, temporal video grounding, moment localization, video-text retrieval, generative augmentation, Bi-Mamba, SIGIR 2026.