Results

Human body pose, shape, and body parts

 

Generalizable multi-view 3D human pose estimation

  • A novel approach for 3D human pose estimation from multiple views that is not limited only to training dataset camera arrangement as previous methods
  • The model achieves state-of-the-art performance when "transferred" between different multi-view datasets
  • The paper presented at CVPR 2022 (New Orleans)
  • Public code available on GitHub

 

 

 

The linear regression baseline for human body measurement

  • A simple baseline takes only a person's height and mass as input and estimates 15 standard body measurements
  • The baseline outperforms most of the existing state-of-the-art, deep learning-based methods
  • The corresponding paper was accepted in Sensors journal in 2022 (Special Issue "Deep Learning and Sensing Technology for Anthropometry")
  • Public code available on GitHub

 

 

 

On using PointNet Architecture for Human Body Segmentation

  • modification of the PointNet architecture for the purpose of human body parts' segmentation
  • Paper presented at the conference ISPA 2019 in Dubrovnik, Croatia
  • Paper available on IEEE Xplore pages

 

 

 

Smartphone solutions for human body analysis

 

A prototype of a catadioptric (multi-mirror) adapter smartphone stereo

  • A 3D-printed prototype based on a novel catadioptric stereo design
  • The adapter provides several degrees of freedom which is particularly important in the prototyping stage
  • The stereo was used for estimating human body measurements from images
  • The corresponding paper is presented at ISPA conference 2021 (virtual)

 

 

 

Smartphone application for body measurements from height and weight

  • A simple Android-based body measurement application
  • Expects human body weight and height as input and shows body shape estimation, such as the one below
  • Public code available on GitHub

 

 

 

Results in 3D point cloud registration

 

Generalizable 3D point cloud registration baseline

  • A baseline approach for 3D point cloud registration which targets particularly the generalizability problem of the previous methods
  • The algorithm achieves state-of-the-art performance on some of the publicly available datasets
  • Corresponding paper presented at the BMVC Workshop 2022
  • Public code available on GitHub.

 

 

 

3D point cloud registration benchmark

  • The evaluation of six selected state-of-the-art registration methods
  • Using Stanford3D public dataset
  • Public code available on GitHub

 

Registration results using: FPFH+RANSAC, FPFH+RANSAC+point-to-point ICP, FPFH+RANSAC+point-to-plane ICP (first row), FGR, FGR+point-to-point ICP, Go-ICP (second row), respectively


 

Other results (image keypoints)

 

On the Comparison of Classic and Deep Keypoint Detector and Descriptor Methods

  • Exhaustive comparison of classical and deep 2D keypoint algorithms
  • Definition of a new detector + descriptor evaluation benchmark
  • Paper presented at the conference ISPA 2019 in Dubrovnik, Croatia
  • Public code available on GitHub.
  • Paper available on IEEE Xplore pages