Awesome

AID-AppEAL: Automatic Image Dataset and Algorithm for Content Appeal Enhancement and Assessment Labeling (ECCV 2024)

[Arxiv] [Poster] [Youtube]

AID-AppEAL is a system that automates dataset creation and implements algorithms to estimate and boost content appeal.

Datasets

Create a dataset root directory called datasets/ in this directory. Download datasets/<dataset_name>.zip from here, move them to the dataset root directory, and unzip them.

For more details of how these datasets are created, see DATASET_CREATION.md

Training

Relative Content Appeal Score Comparator

To train the relative content appeal score comparator on the synthetic dataset for real image dataset labelling, run this command:

bash scripts/1_relative_appeal_score_comparison.sh <dataset_name>

like the following:

bash scripts/1_relative_appeal_score_comparison.sh food
bash scripts/1_relative_appeal_score_comparison.sh room

We provide the checkpoints here under ckpts following the nameing convension ckpts/pair_with_clip_<dataset_name>/last-v1.zip.

Absolute Content Appeal Score Predictor

To train the absolute content appeal score predictor on the real image dataset, run this command:

bash scripts/2_appeal_score_prediction.sh <dataset_name>

We provide the checkpoints here under ckpts following the nameing convension ckpts/singular_with_clip_<dataset_name>/last-v1.zip.

Inference

Content Appeal Estimation and Heatmap Generation

To estimate the content appeal score without generating the heatmap, run this command:

python appeal_heatmap_generation.py --name singular_with_clip_<dataset_name> --input_dir <path_to_input_images> 

To estimate the content appeal score and generate the heatmap, run this command:

python appeal_heatmap_generation.py --name singular_with_clip_<dataset_name> --input_dir <path_to_input_images> --get_appeal_heatmap

Results will be saved under outputs by default.

Content Appeal Enhancement

To enhance image content appeal, we use Automatic1111 Stable Diffusion web UI > img2img > Generation > Inpaint upload with Stable Diffusion v2.1, where aforementioned heatmaps are used as inpainting masks. For more details, pleae refer to the paper and the supplementary pdf.

BibTeX

@misc{chen2024aidappealautomaticimagedataset,
      title={AID-AppEAL: Automatic Image Dataset and Algorithm for Content Appeal Enhancement and Assessment Labeling}, 
      author={Sherry X. Chen and Yaron Vaxman and Elad Ben Baruch and David Asulin and Aviad Moreshet and Misha Sra and Pradeep Sen},
      year={2024},
      eprint={2407.05546},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2407.05546}, 
}