Offical PyTorch implementation of the paper Target-absent Human Attention (ECCV2022)
The prediction of human gaze behavior is important for building human-computer interactive systems that can anticipate a user's attention. Computer vision models have been developed to predict the fixations made by people as they search for target objects. But what about when the image has no target? Equally important is to know how people search when they cannot find a target, and when they would stop searching. In this paper, we propose the first data-driven computational model that addresses the search-termination problem and predicts the scanpath of search fixations made by people searching for targets that do not appear in images. We model visual search as an imitation learning problem and represent the internal knowledge that the viewer acquires through fixations using a novel state representation that we call Foveated Feature Maps (FFMs). FFMs integrate a simulated foveated retina into a pretrained ConvNet that produces an in-network feature pyramid, all with minimal computational overhead. Our method integrates FFMs as the state representation in inverse reinforcement learning. Experimentally, we improve the state of the art in predicting human target-absent search behavior on the COCO-Search18 dataset
If you are using this work, please cite:
@InProceedings{Yang_2022_ECCV_target,
author = {Yang, Zhibo, Sounak Mondal, Seoyoung Ahn, Gregory Zelinsky, Minh Hoai, and Dimitris Samaras},
title = {Target-absent Human Attention},
booktitle = {The European conference on computer vision (ECCV)},
month = {October},
year = {2022}
}conda create --name my_env python=3.8 -y
conda activate my_env
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch
pip install -U opencv-python
git clone git@github.com:cvlab-stonybrook/Target-absent-Human-Attention.git
cd Target-absent-Human-Attention
pip install -r requirements.txt- Train a model with
python train.py --hparams ./configs/coco_search18_TA_IQL.json --dataset-root <dataset-path> - Train termination predictor (set the checkpoint path in hparams.Model.checkpoint first)
python train_terminator_and_eval.py --hparams ./configs/coco_search18_TA_IQL_stop.json --dataset-root <dataset-path> - Evaluation
python train.py --hparams <hparams-path> --dataset-root <dataset-path> --eval-only
We follow the settings in IRL (Yang et al., CVPR2020) and use an action space of 20x32 and COCO-Search18 as the training and evaluation dataset.