This repository contains the tensorflow implementation of paper "Predicting visual attention and distraction during visual search using convolutional neural networks"
Abstract: Most studies in computational modeling of visual attention are focused on task-free observation of images. However, free-viewing saliency considers very limited scenarios of humans’ behavior in daily life. Most visual activities that humans perform daily are goal-oriented and demand a great amount of top-down attention control. Visual Search is a simple task that requires observers to look for a target object in a scene. Compared to free-viewing saliency, visual search demands more top-down control of attention. In this paper, we present two approaches to model visual attention and distraction of observers during visual search. In the first approach, we adapt a light-weight free-viewing saliency model to predict fixation density maps of human observers as the probability of eye fixation over pixels of the search image, using a two-stream encoder-decoder network. This method helps us predict which locations are more distracting when searching for a particular target. We use the COCO-Search18 dataset to train and evaluate our model. Our network achieves good results on the standard saliency metrics (AUC-Judd=0.95, AUC-Borji=0.85, sAUC=0.84, NSS=4.64, KLD=0.93, CC=0.72, SIM=0.54, and IG=2.59). Our second approach is object-based and predicts the location of distractor and target objects during visual search. Distractors are defined as all objects other than the target that the observers fixate on during their search. This method uses a Mask-RCNN segmentation network pre-trained on MS-COCO and fine-tuned on the COCO-Search18 dataset. We release our segmentation annotations of targets and distractors in images of COCO-Search18 for three target object categories: bottle, bowl, and car. For each target object category, we train and evaluate a separate model. The average scores over the three categories are: F1-score=0.64, MAP0.5=0.57, MAR0.5=0.73.
!!ATTENTION!!: Our second method's implementation code and README can be found in 👉 target-distractor_segmentation folder.
Original repository forked from the implementation of MSI-Net saliency network Contextual encoder-decoder network for visual saliency prediction (2020), original implementation
If you use the code of this repository, please do not forget to give us a star 😇 and cite the following:
@misc{https://doi.org/10.48550/arxiv.2210.15093,
doi = {10.48550/ARXIV.2210.15093},
url = {https://arxiv.org/abs/2210.15093},
author = {Samiei, Manoosh and Clark, James J.},
keywords = {Computer Vision and Pattern Recognition (cs.CV), Artificial Intelligence (cs.AI), Machine Learning (cs.LG), Image and Video Processing (eess.IV), FOS: Computer and information sciences, FOS: Computer and information sciences, FOS: Electrical engineering, electronic engineering, information engineering, FOS: Electrical engineering, electronic engineering, information engineering, I.2.10; I.4.9; I.4.6; I.4.m},
title = {Predicting Visual Attention and Distraction During Visual Search Using Convolutional Neural Networks},
publisher = {arXiv},
year = {2022},
copyright = {Creative Commons Attribution 4.0 International}
}
| Package | Version |
|---|---|
| python | 3.7.6 |
| tensorflow | 1.15.0 |
| matplotlib | 3.1.3 |
| numpy | 1.18.1 |
| cv2 | 4.3.0 |
| pandas | 1.0.1 |
| gdown | 4.4.0 |
| wget | 3.2 |
| pysaliency | - |
All dependencies can be installed in a single docker image or an environment.
dataset_download.py downloads COCO-Search18 dataset, target images, target bounding boxes, and VGG16 pretrained weights on ImageNet.
data_preprocessing.py creates task-image pairs. Processes fixation data and creates Gaussian-blurred fixation maps. It resizes all images and fixation maps, augments data with horizontal flips, splits augmented data into train-test-validation sets. Unblurred fixation maps are also generated for test split to be used in saliency metrics computation.
config.py contains the hyperparameters of training such as batch size and number of epochs.
data.py prepares and fetches the data to be efficiently used by gpu. It also handles the iteration over data during training and testing.
model.py defines the model architecture, loads the pretrained weights, and performs the optimization and training operation.
loss.py contains the KLD loss function, that is computed between the groundtruth and predicted saliecy maps.
utils.py represents the training history of a model, keeps track of training and validation error, visualizes the training progress using a progress bar and prints the summary of training and validation loss.
main.py defines the paths to data, results and weights, and handles the training and testing procedures by calling other scripts.
compute_saliency_metrics.py computes the saliency metrics for each image of the test set by calling metrics.py script. It is possible to also use pysaliency library for measuring the metrics. However for consistency with the paper results, you should use metrics.py. The average metric results (of metrics.py) are written in a csv file.
metrics.py contains the the implementation of all saliency metrics namely: AUC-Judd, AUC-Borji, sAUC, NSS, KLD, CC, SIM, and IG.
Run run_all.sh to run all steps including:
1- downloading the dataset
2- data preprocessing
3- training
4- testing
5- computing saliency metrics
You need to change the arguments related to each script accordingly. In the following, we explain each step in details.
Run dataset_download.py as below:
python dataset_download.py \
--dldir=$DOWNLOAD_DIR
dldir: is the directory path to save the downloaded dataset and weights. The default vlues is '../'.
Run data_preprocessing.py as:
python data_preprocessing.py \
--dldir=$DOWNLOAD_DIR \
--sigma=$SIGMA
dldir is the directory path to save the downloaded dataset and weights. The default vlues is '../'.
sigma determines the Gaussian blurring standard deviation. The default value of 11 is used. As in MIT saliency benchmark we set the cut-off frequency f_c as 8. Using formula we derive a sigma of
10.31 and round it up to 11.
To train the model on the dataset, run main.py as below:
python main.py \
--path=$DOWNLOAD_DIR \
--phase=$PHASE \
--threshold=$THRESHOLD
path is the path to the downloaded dataset and weights. The default vlues is '../'.
phase defines whether to train/test the model on the data. The default vlues is 'train'.
threshold defines the threshold for creating binary mask used during jet-color heatmap generation. The default vlues is 30.
To test the model on the dataset, we should change the phase to 'test' and re-run main.py:
python main.py \
--path=$DOWNLOAD_DIR \
--phase=$PHASE \
--threshold=$THRESHOLD
path is the path to the downloaded dataset and weights. The default vlues is '../'.
phase defines whether to train/test the model on the data. The default vlues is 'train'.
threshold defines the threshold for creating binary mask used during jet-color heatmap generation. The default vlues is 30.
To compute saliency metrics, run compute_saliency_metrics.py as:
python compute_saliency_metrics.py \
--path=$DOWNLOAD_DIR \
--use-pysaliency=$PYSAL \
--csv-path=$CSV_DIR
path is the path to the downloaded dataset and weights. The default vlues is '../'.
use-pysaliency is the flag that specifies whether to use pysaliency library to compute saliency metrics. The default vlues is False.
csv-path specifies the path to save the csv file containing computed saliency metrics. The default vlues is '../results/'.
