Usage

May 6, 2026 · View on GitHub

1. Environment Setup

Create the reproduction environment using the following command:

conda env create -f environment.yml

2. Data Preparation

Target Dataset

POPE (COCO split)

Put the dataset under /your_path/data and replace the placeholder in the code with your actual dataset path to ensure proper loading.

3. Usage Instructions

  1. Navigate to the project directory:
cd ./OTT
  1. Full-process inference commands (run sequentially):
# First inference script
bash run_pope.sh

# Second inference script
bash test_pope_my.sh

4. Key Notes

  • This repository provides code for LLaVA-1.5 on the POPE COCO dataset

  • Core logic of the proposed framework implemented in pope_fast.py.

  • Follow the placeholders in the code to configure dataset paths and model loading parameters.

  • This code draws heavily on the open-source work of PAI 、VISTA and MemVR, and we sincerely thank their contributions to the community.