GOAL⚽: Global-local Object Alignment Learning

Implement of paper: GOAL: Global-local Object Alignment Learning

🔍 Project Page

Visit our project page for additional information and interactive examples:

• https://perceptualai-lab.github.io/GOAL/

🐳 Docker

Our implementation is also available as a Docker image:

• Docker Hub

# Pull the image
docker pull qkenr0804/goal:goal

🏋️ Fine-tuned Weights

Download our fine-tuned weights from the links below:

• 🔍 ViT-Base16 Model: GOAL method fine-tuned with DOCCI

  • Download link

• 🔍 ViT-Base16 Model: GOAL method fine-tuned with DCI

  • Download link

• 🔍 ViT-Large14 Model: GOAL method fine-tuned with DOCCI

  • Download link

• 🔍 ViT-Large14 Model: GOAL method fine-tuned with DCI

  • Download link

📊 Datasets

Please download the datasets from the links below:

• DOCCI Dataset

  • Download link

• DCI Dataset

  • Download link

For our newly proposed evaluation protocols on DCI test set and ShareGPT4V test set, please refer to the JSON files available in the datasets folder of this repository.

🚀 Training

You can fine-tuning the CLIP with GOAL method in goal_loss_finetuning.py

You can adjust datasets, ouput path, ... in get_args_parser()

python goal_loss_finetuning.py

📈 Evaluatation

Use your fine-tunned weight

You can evaluate retreival score using retrival_goal.py

You can evaluate mAP score about global+local test set using mAP_goal_jointtset.py

python retrieval_goal.py --ckpt <path/to/your/weight>

python mAP_goal_jointtest.py --ckpt <path/to/your/weight>

👁️ Visualization

You can extract the attention map with you custum weight using visualization_attentionmap.py

python visualization_attentionmap.py --image_path <path/to/your/image> --output_path <path/to/your/output> --model L --ckpt <path/to/your/weight>

📝 Citation

If you find this work useful in your research, please consider citing:

@inproceedings{Hyungyu_2025_CVPR,
  author={Hyungyu Choi, Young Kyun Jang, Chanho Eom},
  title={GOAL: Global-local Object Alignment Learning},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2025}
}