Skip to content

Where do people look on images in average? At rare, thus surprising things! Let's compute them automatically

Notifications You must be signed in to change notification settings

numediart/VisualAttention-RareFamily

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

16 Commits
 
 
 
 
 
 

Repository files navigation

VisualAttention-RareFamily

We provide several codes to compute image saliency from the Rare family. The philosophy of those models is that a specific feature does not necessarily attract human attention, but what attracts it is a feature which is rare, thus surprising and difficult to learn.

Rariy & Saliency Initial image on the left and raw saliency map (probability for each pixel to attract human attention) on the right. No filtering or centred Gaussian applied here.

DeepRare2019 - (DR2019)

Rarity is computed on the deep features extracted by a VGG16 trained on ImageNET. No training is needed. This model is neither "feature-engineered saliency model" as features come from a DNN model, nor a DNN-based model as it needs no training on an eye-tracking dataset: the default ImageNET training of the provided VGG16 is used. It is thus a "deep-engineered" model.

Use DR2019

A full paper can be found here : https://arxiv.org/abs/2005.12073 and here is the Github Project page .

Cite DR2019

If you use DR2019, please cite :

@misc{matei2020visual, title={Visual Attention: Deep Rare Features}, author={Mancas Matei and Kong Phutphalla and Gosselin Bernard}, year={2020}, eprint={2005.12073}, archivePrefix={arXiv}, primaryClass={cs.CV}}

Special strength of DR2019

  • Fully generic model with no training needed. Just run it on your images!
  • Works better than Rare2012 and any other feature-engieneered model and better than some DNN-based models on general images dataset (MIT, ...)
  • Works better than any DNN-based model on one-odd-out datasets (like P3, O3, ...) and is always in top-3 withe feature-engineered models
  • Let you check the contributions of different VGG16 layers to the final result
  • Fast even when ran only on CPU
  • Interesting also for compression applications as the saliency map is precise

Rare 2012 - (R2012)

Rarity is computed on 1) color and 2) Gabor features. This model is a "feature-engineered saliency model".

Use R2012

A full paper can be found here : Main Rare2012 paper and here is the Github Project page .

Cite R2012

If you use R2012, please cite :

@article{riche2013rare2012, title={Rare2012: A multi-scale rarity-based saliency detection with its comparative statistical analysis}, author={Riche, Nicolas and Mancas, Matei and Duvinage, Matthieu and Mibulumukini, Makiese and Gosselin, Bernard and Dutoit, Thierry}, journal={Signal Processing: Image Communication}, volume={28}, number={6}, pages={642--658}, year={2013}, publisher={Elsevier} }

Special strength of R2012

  • Generic ans easy to use
  • Better than R2007

Rare 2007 - (R2007)

Rarity is computed only on color features. This model is a "feature-engineered saliency model".

Use R2007

A full paper can be found here : Main Rare2007 paper and here is the Github Project page .

Cite R2007

If you use R2007, please cite :

@inproceedings{mancas2008relative, title={Relative influence of bottom-up and top-down attention}, author={Mancas, Matei}, booktitle={International Workshop on Attention in Cognitive Systems}, pages={212--226}, year={2008}, organization={Springer} }

Special strength of R2007

  • Generic ans easy to use
  • Interesting for compression applications as it provides a precise saliency map