HHI caption summarization model
This is the summarization model from "Learning Human-Human Interactions in Images from Weak Textual Supervision" (ICCV 2023): a T5-base model fine-tuned to summarize captions into short human-human interaction (HHI) descriptions. It is used to generate the pseudo-labels (pHHI) for the Who's Waldo dataset used to train the main HHI understanding model.
- Paper: arXiv:2304.14104
- Code: github.com/tau-vailab/learning-interactions
- Project page: https://learning-interactions.github.io/
Training data
Fine-tuned on synthetic caption data (synthetic_captions.csv, available in the GitHub repo), mapping full captions to their corresponding HHI descriptions.
Usage
Can be loaded directly with transformers:
from transformers import pipeline
pipe = pipeline('summarization', model='malper/learning-interactions-summarization', device=0)
pipe('summarize: ' + caption)
Or used with the pseudo-labeling code in the repo above (pseudo-labeling/create_pseudolabels.py, pass via -m malper/learning-interactions-summarization or after downloading locally with hf download malper/learning-interactions-summarization --local-dir output/summarization_model).
Training hyperparameters
- learning_rate: 5e-05
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 3.0
Framework versions
- Transformers 4.18.0
- PyTorch 1.13.0a0+d0d6b1f
- Datasets 2.8.0
- Tokenizers 0.12.1
Context
This is research code from 2023, prior to the widespread availability of general-purpose vision-language models (VLMs). It is provided as-is for reproducibility of the paper's results.
Citation
@InProceedings{alper2023learning,
author = {Morris Alper and Hadar Averbuch-Elor},
title = {Learning Human-Human Interactions in Images from Weak Textual Supervision},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
year = {2023}
}