README.md

Composition and Content-Based Image Retrieval(CCBIR)

This is the official webpage of the paper "Enhancing Historical Image Retrieval with Compositional Cues", accepted to AIRoV – The First Austrian Symposium on AI, Robotics, and Vision

Arxiv: https://arxiv.org/abs/2403.14287

News:

• Mar. 27, 2024: We delivered an oral presentation of this paper at AIROV.
• Mar. 20, 2024: We release the code for CCBIR.

1 Network Architectures

2 Datasets

Our work is divided into two main segments: the task of image composition classification and the task of image retrieval.

The dataset employed for the image composition classification task is the KU-PCP dataset, introduced by J.-T. Lee et al.

For the image retrieval task, we extract data from the HISTORIAN dataset, which is richly annotated. The extraction process is shown as follows, and for more details, please refer to our paper:

3 Results

Visualisation of KCM effect. The top row features the original grayscale images, and the bottom row highlights the KCM, pinpointing key compositional areas as detected by our model:

Scatter plot and histogram of positive and negative samples when $L_{KCM}$ is 0.5:

Comparison of retrieval results with different $L_{KCM}$. We selected only the central frame image from each shot in the test set as the target database for retrieval, returning the five highest similarity-scored images for a single image query:

4 Implementation

The implementation of the CAM and KCM mechanisms within CCNet draws significantly from CACNet-Pytorch.

4.1 File Structure

CCBIR:.
|   
+---CBIRNet
|       dataset.py
|       network.py
|       network_cc.py
|       network_module.py
|       select_frames.py
|       test.py
|       test_one.py
|       test_vis.py
|       train.py
|       trainer.py
|       transform.py
|       utils.py
|       validation.py
|       __init__.py
|       
+---CCNet
|       dataset.py
|       network.py
|       network_module.py
|       train.py
|       trainer.py
|       transform.py
|       utils.py
|       validation.py
|       __init__.py

4.2 Required Libraries

Please refer to environment.yml. Create a new conda environment using the following command:

conda env create -f environment.yml

4.3 Training & Testing

During the training phase:

Please ensure that the required files are placed in the data folder, then run train.py to commence training.

During the testing phase:

For CCNet, please run validation.py to test your model. Download the pre-trained CCNet model from this link.

For CBIRNet, use test.py to conduct an overall performance evaluation on a test set; employ test_vis.py for visualization of the results generated by test.py; utilize test_one.py to find the closest image in a specified image folder for a given input picture.

4.4 Build Your Own CBIRNet Dataset

Please use select_frames.py to build your own dataset.

5 Citation

If you find this work useful for your research, please cite arxiv version:

@article{lin2024enhancing,
  title={Enhancing Historical Image Retrieval with Compositional Cues},
  author={Lin, Tingyu and Sablatnig, Robert},
  journal={arXiv preprint arXiv:2403.14287},
  year={2024}
}

Please create an issue or contact tylin@cvl.tuwien.ac.at for further questions.