About me

Unsupervised 3D Pose Representation

Most recent view-invariant action recognition and performance assessment approaches rely on a large amount of annotated 3D skeleton data to extract view-invariant features. However, acquiring 3D skeleton data can be cumbersome, if not impractical, in in-the-wild scenarios. To overcome this problem, we present a novel unsupervised approach that learns to extract view-invariant 3D human pose representation from a 2D image without using 3D joint data. Our model is trained by exploiting the intrinsic view-invariant properties of human pose between simultaneous frames from different viewpoints and their equivariant properties between augmented frames from the same viewpoint. We evaluate the learned view-invariant pose representations for two downstream tasks. We perform comparative experiments that show improvements on the state-of-the-art unsupervised cross-view action classification accuracy on NTU RGB+D by a significant margin, on both RGB and depth images. We also show the efficiency of transferring the learned representations from NTU RGB+D to obtain the first ever unsupervised cross-view and cross-subject rank correlation results on the multi-view human movement quality dataset, QMAR, and marginally improve on the-state-of-the-art supervised results for this dataset. We also carry out ablation studies to examine the contributions of the different components of our proposed network.

VI-Net

We propose a view-invariant method towards the assessment of the quality of human movements which does not rely on skeleton data. Our end-to-end convolutional neural network consists of two stages, where at first a view-invariant trajectory descriptor for each body joint is generated from RGB images, and then the collection of trajectories for all joints are processed by an adapted, pretrained 2D CNN (e.g. VGG-19 or ResNeXt-50) to learn the relationship amongst the different body parts and deliver a score for the movement quality. We release the only publicly-available, multiview, non-skeleton, non-mocap, rehabilitation movement dataset (QMAR), and provide results for both cross-subject and cross-view s enarios on this dataset. We show that VI-Net achieves average rank correlation of 0.66 on cross-subject and 0.65 on unseen views when trained on only two views. We also evaluate the proposed method on the single-view rehabilitation dataset KIMORE and obtain 0.66 rank correlation against a baseline of 0.62.

QMAR is an RGB multi-view dataset for healthcare applications. It was recorded using 6 Primesense cameras with 38 healthy subjects, 8 female and 30 male. The subjects were trained by a physiotherapist to perform two different types of movements while simulating two ailments, resulting in four overall possibilities: a return walk to approximately the original position while simulatin Parkinsons (W-P), and Stroke (WS), and standing up and sitting down with Parkinson (SS-P) and Stroke (SS-S). The dataset includes RGB and depth (and skeleton data, although in this work we only use RGB. As capturing depth data from the 6 Primesense cameras was not possible due to infrared interference, the depth and skeleton data were retaine from only view 2 at ≈ 0◦ and view 5 at ≈ 90◦. The movements in QMAR were scored by the severity of the abnormality. The score ranges were 0 to 4 for W-P, 0 to 5 for W-S and SS-S, and 0 to 12 for SS-P.

View-Invariant Human Pose

We propose a CNN regression method to generate high-level, view-invariant features from RGB images which are suitable for human pose estimation and movement quality analysis. The inputs to our network are body joint heatmaps and limb-maps to help our network exploit geometric relationships between different body parts to estimate the features more accurately. A new multiview and multimodal human movement dataset is also introduced part of which is used to evaluate the results of the proposed method. We present comparative experimental results on pose estimation using a manifold-based pose representation built from motion-captured data. We show that the new RGB derived features provide pose estimates of similar or better accuracy than those produced from depth data, even from single views only.

2021

F. Sardari, B. Ommer, and M. Mirmehdi, "Unsupervised View-Invariant Human Posture Representation", British Machine Vision Conference (BMVC), 2021.

2020

F. Sardari, A. Paiement, S. Hannuna, and M. Mirmehdi, "VI-Net: View-Invariant Quality of Human Movement Assessment", Sensors Journal, 2020.

2019

F. Sardari, A. Paiement, and M. Mirmehdi, "View-invariant Pose Analysis for Human Movement Assessment from RGB Data", International Conference on Image Analysis and Processing (ICIAP), 2019.

2017

F. Sardari, and M. E. Moghaddam, "A hybrid occlusion free object tracking method using particle filter and modified galaxy based search meta-heuristic algorithm", Applied Soft Computing Journal, 2017.

2016

F. Sardari, and M. E. Moghaddam, "An object tracking method using modified galaxy-based search algorithm", Swarm and Evolutionary Computation Journal, 2016.

2012

F. Sardari, and M. E. Moghaddam, "A genetic based generic filter for image impulse noise reduction", IET Image Processing Conference, 2012.

Office

Visual Information Laboratory, University of Bristol

1 Cathedral Square

Trinity Street

Bristol, BS1 5DD

E-mail

faegheh.sardari(at)bristol.ac.uk