AI Can Spot Objects Even If They Are Hidden

CAMOUFLAGED objects are  difficult to detect, for both  humans and artificial  intelligence. But now an AI has  been trained to parse objects  from their backgrounds.

AI


This could have a variety  of applications, such as being  used for search-and-rescue  work, detecting agricultural  pests, medical imaging or in  military settings.

Detecting camouflaged objects  requires visual perception and  knowledge. Until now, many AIs have struggled with this task  because their algorithms rely on  visual cues, such as differences   in colour or easily recognisable  shapes, to identify objects.

To improve on this, Jianbing  Shen at the Inception Institute of  Artificial Intelligence in Abu Dhabi  in the United Arab Emirates and  his colleagues collated a data set   of 10,000 photographs to train   an AI. The data set includes 5066  images of camouflaged objects,  which they have divided into 78  categories, such as “amphibian”,  “aquatic” and “flying”.

The photographs included  both naturally camouflaged  animals such as fish and insects  and examples of artificial  camouflage, such as soldiers  in uniform. Although databases  of camouflaged objects already  exist, this data set is the largest,  says Shen.

The team manually labelled  each image of a camouflaged  object to highlight characteristics  such as its shape or whether it  was partially obstructed by its  surrounding environment.  

They then developed an AI  called SINet and trained it on  images from the data set.  The researchers compared   SINet to 12 existing algorithms  built to detect generic objects. They tested all 13 algorithms  using three existing data sets  of camouflaged objects. SINet

“ Many AIs struggle to detect  camouflaged objects  because their algorithms  rely on visual cues”

did better than the other 12 at   isolating camouflaged objects and  identifying their correct shape   and nature in both the existing  and the training data sets.
“Without any bells and whistles,  SINet outperforms various stateof-the-art object detection  baselines on all datasets tested,  making it a robust, general  framework that can help facilitate  future research,” the researchers  write.  They are due to present the  work at the CVPR 2020 conference  in Seattle, Washington, in June.
The researchers hope the data  set and algorithm can improve AI’s  ability to recognise camouflaged  objects, says Shen.  ❚

See this video of ibm giant company.


Also Read Our latest post :

Here  10 things Nobody tell you about.

The Guide Of Design Thinking To Become Innovative.

/* If you want to remove footer link visit http://shubhammali.cf/ here and contact me bruus */