SentiSight Embedded SDK

Object recognition for moble computer vision applications

SentiSight Embedded is designed for developers who want to use computer vision-based object recognition in their applications for smartphones, tablets and other mobile devices. Through manual or fully automatic object learning it enables searching for learned objects in images or videos from built-in cameras with PC-like accuracy.

Available as a software development kit that provides for the development of object recognition applications for the devices that are running Android OS.

Technical Specifications

A Java application based on SentiSight Embedded 1.3 technology is able to process a image with an object in less than 1 second. Object recognition algorithms can be run on more than one thread of multi-core processors.

These conditions may alter the performance of the algorithms:

  • Rotation and translation. The algorithm is generally rotation- and translation-invariant in a plane perpendicular to the camera. The algorithm is also invariant for rotations of up to 10-15 degrees beyond a plane perpendicular to the camera. Different views of an object may be added to a model to handle larger rotations.
  • Resolution and scale changes. Scale (size of the object within the image) difference between an object's model and the object itself can be up to 2-3 times. Objects should contain enough detail, and be large enough to be recognizable.
  • Occlusions. The algorithm is robust to occlusions as big as 50 % of the object if enough unique edges remain visible.
  • Lighting conditions (illumination, shadows and reflectance).
    • Planar objects will have issues with reflectance.
    • 3D objects will have issues with variable lighting conditions – consistent lighting greatly reduces potential problems.
  • Transparency. In general, transparent objects are difficult to recognize.
  • Rigidity. The algorithm recognizes only rigid/stable objects. At minimum, a significant piece of the object should be unchanging.

Object model size depends on how feature-rich is an object, and thus is individual for each object. All object models should be loaded into RAM before identification, thus the maximum object model database size is limited by the amount of available RAM.

The SentiSight Embedded 1.3 algorithm was tested with a subset of Amsterdam Library of Object Images (ALOI).

  • The subset contained objects 1-100 from ALOI.
  • Images with object viewpoint variations (ALOI-VIEW collection) were used. 36 images per object were used.

The average object model size when testing with the 768 x 576 pixels images (the original resolution) was:

  • 0.7 Megabytes when blob-based algorithm was used.
  • 3.2 Megabytes when shape-based algorithm was used.

When the images were resized to 320 x 240 pixels, the average object model size was:

  • 0.2 Megabytes when blob-based algorithm was used.
  • 0.5 Megabytes when shape-based algorithm was used.

At 0.1 % False Acceptance Rate (FAR), the recognition rate is from 70 % - 99 %, depending on object structural appearance, transparency, etc. For objects with well-defined internal structure, the recognition rate is 98 % - 99 % at 0.1 % FAR.