🎬 (AAAI 2026) ReTrack: Evidence-Driven Dual-Stream Directional Anchor Calibration Network for Composed Video Retrieval (Model Weights)
1School of Software, Shandong University
2School of Computer Science and Technology, Shandong Jianzhu University
✉ Corresponding author
This repository hosts the official pre-trained model weights for ReTrack, an evidence-driven framework designed to calibrate directional bias in composed features for both Composed Video Retrieval (CVR) and Composed Image Retrieval (CIR) tasks.
📌 Model Information
1. Model Name
ReTrack (Evidence-Driven Dual-Stream Directional Anchor Calibration Network) Checkpoints.
2. Task Type & Applicable Tasks
- Task Type: Composed Video Retrieval (CVR) and Composed Image Retrieval (CIR).
- Applicable Tasks: Retrieving a target video or image based on a reference visual input combined with a modification text prompt. The model significantly reduces uncertainty caused by highly similar retrieval candidates in multi-modal queries.
3. Project Introduction
ReTrack is an advanced open-source PyTorch framework built on top of BLIP-2 (via Salesforce LAVIS) that improves multi-modal query understanding. It features:
- 🎯 Dual-Stream Directional Anchor Calibration: Explicitly identifies and calibrates visual and textual semantic contributions to resolve directional bias.
- ⚖️ Reliable Evidence-Driven Alignment: Leverages Dempster-Shafer Theory to evaluate similarity reliability, minimizing ambiguity among candidates.
4. Training Data Source & Hosted Weights
The framework is trained to support both the WebVid-CoVR dataset for video retrieval and the FashionIQ / CIRR datasets for image retrieval.
This Hugging Face repository provides the following pre-trained checkpoint:
- 📄
ReTrack-WebVid-Frame1.ckpt: The checkpoint trained on the WebVid-CoVR dataset (using a 1-frame configuration setting).
🚀 Usage & Basic Inference
These weights are designed to be evaluated using the highly modular, Hydra-configured
ReTrack GitHub repository.
Step 1: Prepare the Environment
We recommend using Anaconda. Clone the repository and install dependencies:
1git clone https://github.com/iLearn-Lab/AAAI26-ReTrack.git
2cd ReTrack
3conda create -n retrack python=3.8 -y
4conda activate retrack
5conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
6pip install -r requirements.txt
Step 2: Download Model Weights & Prepare Data
- Download
ReTrack-WebVid-Frame1.ckpt from this Hugging Face repository.
- Place the checkpoint in the appropriate directory as expected by your Hydra configuration (e.g., within a
checkpoints/ folder).
- Ensure the WebVid-CoVR dataset is placed under your defined
datasets_dir in configs/machine/default.yaml.
Step 3: Run Evaluation
To evaluate the trained CVR model, use test.py and specify the path to your downloaded checkpoint via Hydra CLI overrides:
1python test.py \
2 model.ckpt_path=/path/to/your/ReTrack-WebVid-Frame1.ckpt \
3 +test=webvid-covr
(Refer to the configs/ directory in the code repository for advanced hyperparameter and path adjustments).
⚠️ Limitations & Notes
- Configuration: ReTrack is entirely managed by Hydra and Lightning Fabric. Make sure you are familiar with overriding configurations via the CLI or modifying the YAML files in the
configs/ directory.
- Environment: The project was specifically developed and evaluated on Python 3.8 and PyTorch 2.1.0; using drastically different versions may yield unexpected behaviors.
📝⭐️ Citation
If you find our framework, code, or these weights useful in your research, please consider leaving a Star ⭐️ on our GitHub repository and citing our AAAI 2026 paper:
1@inproceedings{ReTrack,
2 title={ReTrack: Evidence Driven Dual Stream Directional Anchor Calibration Network for Composed Video Retrieval},
3 author={Li, Zixu and Hu, Yupeng and Chen, Zhiwei and Huang, Qinlei and Qiu, Guozhi and Fu, Zhiheng and Liu, Meng},
4 booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
5 year={2026}
6}
