Releases: Telecommunication-Telemedia-Assessment/GBVS360-BMS360-ProSal
Build for experimentations
This build enable some of the command line options to check the effect of different parameters in the model.
DO NOT USE FOR BENCHMARKING PURPOSE. USE THE OTHERS RELEASE INSTEAD.
For example, please use this version if your goal is to compare these models with yours: https://github.com/Telecommunication-Telemedia-Assessment/GBVS360-BMS360-ProSal/releases/tag/2.0
This tool was based on the binary submitted to the ICME grand challenges. It allow to test the differences of model performance with/without equatorial prior. This should not be used for benchmarking purpose as the equatorial prior is part of the contribution of these model and disabling it results in decrease of performance in the context of Head Saliency prediction. Therefore, please use the other build of the release section if you aim at doing comparison between this work and yours.
The models available in this binary are: BMS360, BMS, and GBVS360:
BMS360
You can obtain the saliency map from the BMS360 model using the following command:
.\salient.exe -i .\test.jpg -o .\saliency.jpg --general-model 1
.\salient.exe -i .\test.jpg -o .\saliency.jpg --general-model 2
There is no difference between the two command, as the intended difference between the two was the use or not of the equatorial prior. However, in this tool, the equatorial prior is ALWAYS DISABLED. (Again, You should NOT USE THAT BUILD FOR BENCHMARKING)
BMS
You can access to the BMS model using the following command line:
.\salient.exe -i .\test.jpg -o .\saliency.jpg --general-model 1 --legacy-icme-2017
Again, the equatorial prior is DISABLED and that should NOT be done. This is for testing ONLY, not BENCHMARKING.
GBVS360
You can access to the GBVS360 model using the following command line:
.\salient.exe -i .\test.jpg -o .\saliency.jpg --general-model 2 --legacy-icme-2017
Again, the equatorial prior is DISABLED and that should NOT be done. This is for testing ONLY, not BENCHMARKING.
Equatorial prior
And in each case, you can also add the equatorial prior using the option --equatorial-prior
Predictions
This archives contains the prediction results from the models BMS360 and GBVS360 when applied to the ICME 2017 Grand Challenge Salient360! dataset.
Please refer to the paper:
@article{Gutierrez2018Toolbox,
title={Toolbox and Dataset for the Development of Saliency and Scanpath Models for Omnidirectional / 360{$^{\circ}$} Still Images},
author={Guti{\'e}rrez, Jes{\'u}s and David, Erwan and Rai, Yashas and Le Callet, Patrick
},
journal={Signal Processing: Image Communication},
volume={69},
pages={35--42},
year={2018},
publisher={Elsevier}
}
To obtain the ground truth data and evaluation scripts.
BMS360, GBVS360
This version contains improvements over the model submitted at the ICME Grand Challenge. The new model called BMS360 have been introduced and is suggested as the new model to be used for Head motion-based saliency and Head/Eye motion-based saliency.
The former models can still be accessed through the options. Please refer to the option --help to obtain more information.
Usage
The following shows example of usage of the tool to perform the saliency prediction.
BMS360 + Equatorial prior : HEAD Saliency model
The Head saliency model submitted at ICME2018 was using BMS360 (with tuned parameters for omnidirectional images) and the "adaptive equatorial prior". The usage is as follow:
salient.exe -i P10.jpg --general-model 1 -o head_saliency.jpg
BMS360 without the equatorial prior : HEAD+Eye Saliency model
The Head+Eye saliency model in this build is using the BMS360 model without the equatorial prior. This model was NOT submitted to ICME2018 as other model are better suited for the Head+Eye saliency prediction. More details can be found here:
@article{lebreton2018GBVS360, title={GBVS360, BMS360, ProSal: Extending existing saliency prediction models from 2D to omnidirectional images}, author={Lebreton, Pierre and Raake, Alexander}, journal={Signal Processing: Image Communication}, volume={69}, pages={69--78}, year={2018}, publisher={Elsevier} }
Nevertheless, to get the result for the BMS360 model without equatorial prior, please run:
salient.exe -i P10.jpg --general-model 2 -o head_eye_saliency.jpg
Legacy model:
This program also contains the model submitted to the ICME2017 grand challenges and you can find the model "BMS+equatorial prior", and "GBVS360 with equatorial prior" using the following command (please note the use of the option --legacy-icme-2017):
BMS + Equatorial prior
The Head saliency model submitted at ICME2017 was using BMS model and the "adaptive equatorial prior". The usage is as follow:
salient.exe -i P10.jpg --general-model 1 -o saliency.jpg --legacy-icme-2017
GBVS360 + Equatorial prior
The Head+Eye saliency model submitted at ICME2017 was using GBVS360 model and the "adaptive equatorial prior". The usage is as follow:
salient.exe -i P10.jpg --general-model 2 -o saliency.jpg --legacy-icme-2017
Submission to ICME Grand Challenge 2017
This the model as submitted to the ICME Grand Challenge 2017: Salient360!
The binary contains the two models, please refer to the README for more information. The binary also contains a manual, see the option --help.
Usage
The following shows example of usage of the tool to perform the saliency prediction.
BMS + Equatorial prior : HEAD Saliency model
The Head saliency model submitted at ICME2017 was using BMS (with tuned parameters for omnidirectional images) and the "adaptive equatorial prior". The usage is as follow:
salient.exe -i P10.jpg --general-model 1 -o head_saliency.jpg
GBVS360 + Equatorial prior : HEAD+EYE Saliency model
The Head+Eye saliency model submitted at ICME2017 was using GBVS360 model and the "adaptive equatorial prior". The usage is as follow:
salient.exe -i P10.jpg --general-model 2 -o head_eye_saliency.jpg