Kun Wang1 Yupeng Hu1 Zhiran Li1 Hao Liu1 Qianlong Xiang2,3,4 Liqiang Nie2
1Shandong University
2Harbin Institute of Technology
3City University of Hong Kong
4Shenzhen Loop Area Institute
--- ## 📌 Model Information ### 1. Model Name **ViSAGE (Video Saliency with Adaptive Gated Experts)** ### 2. Task Type & Applicable Tasks - **Task Type:** Video Saliency Prediction (VSP) / Computer Vision - **Applicable Tasks:** Robust and adaptive prediction of human visual attention (saliency maps) in dynamic video sequences. ### 3. Project Introduction Video Saliency Prediction requires capturing complex spatio-temporal dynamics and human visual priors. **ViSAGE** tackles this by leveraging a powerful multi-expert ensemble framework. > 💡 **Method Highlight:** The framework consists of a shared **InternVideo2 backbone** adapted via two-stage LoRA fine-tuning, alongside dual specialized experts utilizing Temporal Modulation (for explicit spatial priors) and Multi-Scale Fusion (for adaptive data-driven perception). For robust performance, the **Ensemble Fusion Module** obtains the final prediction by converting the expert outputs to logit space before averaging. ### 4. Training Data Source - Dataset provided by the **NTIRE 2026 Video Saliency Prediction Challenge** (Private Test and Validation sets). --- ## 🚀 Usage & Basic Inference ### Step 1: Prepare the Environment Clone the GitHub repository and set up the Conda environment: ```bash git clone https://github.com/iLearn-Lab/CVPRW26-ViSAGE.git cd ViSAGE conda create -n visage python=3.10 -y conda activate visage pip install -r requirements.txt ``` ### Step 2: Data & Pre-trained Weights Preparation 1. **Challenge Data:** Use the provided scripts to extract frames from the source videos. ```bash python video_to_frames.py ``` 2. **InternVideo2 Backbone:** Download the pre-trained `InternVideo2-Stage2_6B-224p-f4` model from [Hugging Face](https://huggingface.co/OpenGVLab/InternVideo2-Stage2_6B-224p-f4) and clone the `InternVideo` repo. 3. **Paths:** Update the pre-trained weight paths in `Expert1/inference.py` and `Expert2/inference.py` to match your local directory. ### Step 3: Run Inference & Ensemble **1. Inference:** Generate predictions for both experts. ```bash python Expert1/inference.py python Expert2/inference.py ``` **2. Ensemble:** Merge the inference results from Expert 1 and Expert 2 in logit space. ```bash python ensemble.py ``` **3. Format Check & Video Generation:** ```bash python check.py python makevideos.py ``` ### Step 4: Training (Optional) Run the two-stage LoRA fine-tuning pipeline: ```bash python trainnew.py # Stage 1 python trainnew2.py # Stage 2 ``` --- ## ⚠️ Limitations & Notes - The model relies heavily on the InternVideo2 backbone; out-of-memory (OOM) errors may occur on GPUs with less than 24GB VRAM. - This framework and its pre-trained weights are intended for **academic research purposes only**. --- ## 📝⭐️ Citation If you find this project useful for your research, please consider citing: ```bibtex @inproceedings{ntire26visage, title={{ViSAGE @ NTIRE 2026 Challenge on Video Saliency Prediction: Methods and Results}}, author={Wang, Kun and Hu, Yupeng and Li, Zhiran and Liu, Hao and Xiang, Qianlong and Nie, Liqiang}, booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops}, year={2026} } ```