Working for many years in computer vision area we completed a lot of projects and have great hands-on experience in development solutions for various image & video processing problems.

Here are just a few of them.


Near duplicate image retrieval

Realtime search for near duplicate images in database with 100M image.

  • Based on visual words concept.
  • Very compact descriptor. Just 1-2Kb for single image.
  • Realtime search over large collection. Our solution is able to retrieve an image against 100,000,000 collection in sub-seconds.
  • Scalable architecture. Our solution is designed for parallel search on the cluster of machines for handling large image collections.

More..

Near duplicate image retrieval

Analysis of X-Ray images

Sample image is from Wikimedia project (CC-BY-SA-4.0): https://commons.wikimedia.org/wiki/File:Cephalometric_radiograph.JPG

Sample image is from Wikimedia project (CC-BY-SA-4.0): https://commons.wikimedia.org/wiki/File:Cephalometric_radiograph.JPG

Robust and efficient algorithm for processing of  x-ray images, based on deep learning approach. Convolutional Neural Networks (CNN) were trained using TensorFlow framework by collection of 100000+ samples.

The solution has the following features:

  • Detection of skull position and orientation.
  • Skull image quality estimation.
  • Scale detection (using ruler image found in the input image).
  • Tracing of 90+ landmarks.
  • Automatic estimation of confidence scores for detected values (points, scales, etc).

 


Age and gender recognition

Age and gender recognition by face on still images and video.

  • Realtime performance.
  • Gender recognition accuracy is 95% on the FERET database.
  • Age recognition is accurate up to 5 years in 88% test cases on the custom collection of real-world images.
  • Server-side solution for processing live video stream with RESTful API.

Text detection

Realtime text segmentation with irregular fonts and shape in the images and videos.

  • Features: multi-scale orientations field.
  • Classifier: Decision Forest.
  • Performance: 50 ms for regular web image.

 


Augmented reality

Augmented reality with third-party tools Qualcomm Vuforia and ARToolkit.

  • For mobile devices.
  • High-tolerant to various light conditions.
  • Pre-defined set of patterns for recognition.

Pattern matching in mobile devices

Detect and recognize different patterns in mobile devices.

  • Based on SURF features.
  • Support half damaged/visible patterns.
  • Rotation invariant.

Eyes detection and segmentation

Automatic detection of eyes region and precise iris segmentation.

  • Realtime processing.
  • Handling of rotated faces, side view, occlusions by hair and glasses.
  • Works for color and grayscale images.
  • Features: multi-scale orientations fields.
  • Classifier: Decision Forest and Hough Trees.
  • Performance: 96% of images have accuracy more than 95%.

Face montage

Face montage

Automatic seamless inpainting of one face to another head.

  • Automatic facial feature detection.
  • Poisson image editing technique for seamless stitching.
  • iOS and Android client applications.
  • Backend hosted on Amazon EC2 infrastructure.

 


Human detection and tracking

Human detection and tracking

Realtime human detection and tracking in live video stream from fisheye camera.

  • Background and foreground segmentation.
  • Resolving ambiguities of crossing tracks.
  • Re-identification of re-entering humans.
  • Server-side solution with HTTP API.
  • Integration with Age and Gender classifier from side-view camera.

Object tracking

Object tracking

Tracking objects of various nature, like humans, vehicles, faces, etc in video.

  • Handling of shadows and highlights.
  • Background and foreground segmentation.
  • Server-side back-end for multiple cameras.

Logo detection

Logotypes detection

Logotypes detection on still images and video.

  • Realtime processing of video stream.
  • Simultaneous recognition of multiple logotypes.
  • HoG-like features with SVM classifier.
  • Detects logos under heavy perspective distortion.

Face detection for Epiphany multicore architecture

Implementation of LBP face detection for massively parallel multicore architecture.

  • Custom hardware architecture.
  • Works in environment with very limited memory.
  • Take advantage of massively parallel architecture (16+ cores).

More..


Licence plate recognition

Recognition of car licence plate in videos from surveillance cameras

  • Automatic detection, recognition and tracking of license plates in video.
  • Based on HoG-like features and SVM classifier.
  • Compatible with low-resolution monochrome images.