scholarly journals Chained Representation Cycling: Learning to Estimate 3D Human Pose and Shape by Cycling Between Representations

2020 ◽  
Vol 34 (04) ◽  
pp. 5561-5569 ◽  
Author(s):  
Nadine Rueegg ◽  
Christoph Lassner ◽  
Michael Black ◽  
Konrad Schindler ◽  
Nadine Rueegg ◽  
...  
Keyword(s):  
Ground Truth ◽  
Body Part ◽  
Training Data ◽  
Shape Estimation ◽  
Paired Data ◽  
One Step ◽  
Computer Vision Systems ◽  
Image Pixels ◽  
Human Pose ◽  
Object Parameters

The goal of many computer vision systems is to transform image pixels into 3D representations. Recent popular models use neural networks to regress directly from pixels to 3D object parameters. Such an approach works well when supervision is available, but in problems like human pose and shape estimation, it is difficult to obtain natural images with 3D ground truth. To go one step further, we propose a new architecture that facilitates unsupervised, or lightly supervised, learning. The idea is to break the problem into a series of transformations between increasingly abstract representations. Each step involves a cycle designed to be learnable without annotated training data, and the chain of cycles delivers the final solution. Specifically, we use 2D body part segments as an intermediate representation that contains enough information to be lifted to 3D, and at the same time is simple enough to be learned in an unsupervised way. We demonstrate the method by learning 3D human pose and shape from un-paired and un-annotated images. We also explore varying amounts of paired data and show that cycling greatly alleviates the need for paired data. While we present results for modeling humans, our formulation is general and can be applied to other vision problems.

scholarly journals 3D Human Pose Estimation Using Spatio-Temporal Networks with Explicit Occlusion Training

2020 ◽  
Vol 34 (07) ◽  
pp. 10631-10638
Author(s):  
Yu Cheng ◽  
Bo Yang ◽  
Bo Wang ◽  
Robby T. Tan
Keyword(s):  
Pose Estimation ◽  
Ground Truth ◽  
Video Data ◽  
Training Data ◽  
Human Pose Estimation ◽  
Ground Truth Data ◽  
Public Data ◽  
Spatio Temporal ◽  
Human Pose ◽  
3D Human Pose Estimation

Estimating 3D poses from a monocular video is still a challenging task, despite the significant progress that has been made in the recent years. Generally, the performance of existing methods drops when the target person is too small/large, or the motion is too fast/slow relative to the scale and speed of the training data. Moreover, to our knowledge, many of these methods are not designed or trained under severe occlusion explicitly, making their performance on handling occlusion compromised. Addressing these problems, we introduce a spatio-temporal network for robust 3D human pose estimation. As humans in videos may appear in different scales and have various motion speeds, we apply multi-scale spatial features for 2D joints or keypoints prediction in each individual frame, and multi-stride temporal convolutional networks (TCNs) to estimate 3D joints or keypoints. Furthermore, we design a spatio-temporal discriminator based on body structures as well as limb motions to assess whether the predicted pose forms a valid pose and a valid movement. During training, we explicitly mask out some keypoints to simulate various occlusion cases, from minor to severe occlusion, so that our network can learn better and becomes robust to various degrees of occlusion. As there are limited 3D ground truth data, we further utilize 2D video data to inject a semi-supervised learning capability to our network. Experiments on public data sets validate the effectiveness of our method, and our ablation studies show the strengths of our network's individual submodules.

scholarly journals Towards Single 2D Image-Level Self-Supervision for 3D Human Pose and Shape Estimation

2021 ◽  
Vol 11 (20) ◽  
pp. 9724
Author(s):  
Junuk Cha ◽  
Muhammad Saqlain ◽  
Changhwa Lee ◽  
Seongyeong Lee ◽  
Seungeun Lee ◽  
...  
Keyword(s):  
Large Scale ◽  
Three Dimensional ◽  
Ground Truth ◽  
Small Subset ◽  
Learning Approaches ◽  
Shape Estimation ◽  
Camera System ◽  
Ground Truth Data ◽  
Human Pose ◽  
2D Images

Three-dimensional human pose and shape estimation is an important problem in the computer vision community, with numerous applications such as augmented reality, virtual reality, human computer interaction, and so on. However, training accurate 3D human pose and shape estimators based on deep learning approaches requires a large number of images and corresponding 3D ground-truth pose pairs, which are costly to collect. To relieve this constraint, various types of weakly or self-supervised pose estimation approaches have been proposed. Nevertheless, these methods still involve supervision signals, which require effort to collect, such as unpaired large-scale 3D ground truth data, a small subset of 3D labeled data, video priors, and so on. Often, they require installing equipment such as a calibrated multi-camera system to acquire strong multi-view priors. In this paper, we propose a self-supervised learning framework for 3D human pose and shape estimation that does not require other forms of supervision signals while using only single 2D images. Our framework inputs single 2D images, estimates human 3D meshes in the intermediate layers, and is trained to solve four types of self-supervision tasks (i.e., three image manipulation tasks and one neural rendering task) whose ground-truths are all based on the single 2D images themselves. Through experiments, we demonstrate the effectiveness of our approach on 3D human pose benchmark datasets (i.e., Human3.6M, 3DPW, and LSP), where we present the new state-of-the-art among weakly/self-supervised methods.

scholarly journals Quantifying the structure of strong gravitational lens potentials with uncertainty-aware deep neural networks

2020 ◽  
Vol 499 (4) ◽  
pp. 5641-5652
Author(s):  
Georgios Vernardos ◽  
Grigorios Tsagkatakis ◽  
Yannis Pantazis
Keyword(s):  
Confidence Intervals ◽  
Galaxy Evolution ◽  
Gravitational Lensing ◽  
Probability Distributions ◽  
Mass Density ◽  
Ground Truth ◽  
Gaussian Random Fields ◽  
Training Data ◽  
Gravitational Lens ◽  
Data Set

ABSTRACT Gravitational lensing is a powerful tool for constraining substructure in the mass distribution of galaxies, be it from the presence of dark matter sub-haloes or due to physical mechanisms affecting the baryons throughout galaxy evolution. Such substructure is hard to model and is either ignored by traditional, smooth modelling, approaches, or treated as well-localized massive perturbers. In this work, we propose a deep learning approach to quantify the statistical properties of such perturbations directly from images, where only the extended lensed source features within a mask are considered, without the need of any lens modelling. Our training data consist of mock lensed images assuming perturbing Gaussian Random Fields permeating the smooth overall lens potential, and, for the first time, using images of real galaxies as the lensed source. We employ a novel deep neural network that can handle arbitrary uncertainty intervals associated with the training data set labels as input, provides probability distributions as output, and adopts a composite loss function. The method succeeds not only in accurately estimating the actual parameter values, but also reduces the predicted confidence intervals by 10 per cent in an unsupervised manner, i.e. without having access to the actual ground truth values. Our results are invariant to the inherent degeneracy between mass perturbations in the lens and complex brightness profiles for the source. Hence, we can quantitatively and robustly quantify the smoothness of the mass density of thousands of lenses, including confidence intervals, and provide a consistent ranking for follow-up science.

scholarly journals Automated analysis of 3D-echocardiography using spatially registered patient-specific CMR meshes

2021 ◽  
Vol 22 (Supplement_1) ◽  
Author(s):  
D Zhao ◽  
E Ferdian ◽  
GD Maso Talou ◽  
GM Quill ◽  
K Gilbert ◽  
...  
Keyword(s):  
New Zealand ◽  
Interobserver Variability ◽  
Ground Truth ◽  
Automated Analysis ◽  
3D Echocardiography ◽  
Training Data ◽  
Patient Specific ◽  
Manual Analysis ◽  
Lv Mass ◽  
3D Echo

Abstract Funding Acknowledgements Type of funding sources: Public grant(s) – National budget only. Main funding source(s): National Heart Foundation (NHF) of New Zealand Health Research Council (HRC) of New Zealand Artificial intelligence shows considerable promise for automated analysis and interpretation of medical images, particularly in the domain of cardiovascular imaging. While application to cardiac magnetic resonance (CMR) has demonstrated excellent results, automated analysis of 3D echocardiography (3D-echo) remains challenging, due to the lower signal-to-noise ratio (SNR), signal dropout, and greater interobserver variability in manual annotations. As 3D-echo is becoming increasingly widespread, robust analysis methods will substantially benefit patient evaluation.  We sought to leverage the high SNR of CMR to provide training data for a convolutional neural network (CNN) capable of analysing 3D-echo. We imaged 73 participants (53 healthy volunteers, 20 patients with non-ischaemic cardiac disease) under both CMR and 3D-echo (<1 hour between scans). 3D models of the left ventricle (LV) were independently constructed from CMR and 3D-echo, and used to spatially align the image volumes using least squares fitting to a cardiac template. The resultant transformation was used to map the CMR mesh to the 3D-echo image. Alignment of mesh and image was verified through volume slicing and visual inspection (Fig. 1) for 120 paired datasets (including 47 rescans) each at end-diastole and end-systole. 100 datasets (80 for training, 20 for validation) were used to train a shallow CNN for mesh extraction from 3D-echo, optimised with a composite loss function consisting of normalised Euclidian distance (for 290 mesh points) and volume. Data augmentation was applied in the form of rotations and tilts (<15 degrees) about the long axis. The network was tested on the remaining 20 datasets (different participants) of varying image quality (Tab. I). For comparison, corresponding LV measurements from conventional manual analysis of 3D-echo and associated interobserver variability (for two observers) were also estimated. Initial results indicate that the use of embedded CMR meshes as training data for 3D-echo analysis is a promising alternative to manual analysis, with improved accuracy and precision compared with conventional methods. Further optimisations and a larger dataset are expected to improve network performance. (n = 20) LV EDV (ml) LV ESV (ml) LV EF (%) LV mass (g) Ground truth CMR 150.5 ± 29.5 57.9 ± 12.7 61.5 ± 3.4 128.1 ± 29.8 Algorithm error -13.3 ± 15.7 -1.4 ± 7.6 -2.8 ± 5.5 0.1 ± 20.9 Manual error -30.1 ± 21.0 -15.1 ± 12.4 3.0 ± 5.0 Not available Interobserver error 19.1 ± 14.3 14.4 ± 7.6 -6.4 ± 4.8 Not available Tab. 1. LV mass and volume differences (means ± standard deviations) for 20 test cases. Algorithm: CNN – CMR (as ground truth). Abstract Figure. Fig 1. CMR mesh registered to 3D-echo.

scholarly journals ChemProps: A RESTful API enabled database for composite polymer name standardization

2021 ◽  
Vol 13 (1) ◽  
Author(s):  
Bingyin Hu ◽  
Anqi Lin ◽  
L. Catherine Brinson
Keyword(s):  
Data Exchange ◽  
Ground Truth ◽  
Training Data ◽  
Test Accuracy ◽  
Polymer Science ◽  
Data Systems ◽  
Materials Informatics ◽  
Related Data ◽  
Weight Factors ◽  
Data Points

AbstractThe inconsistency of polymer indexing caused by the lack of uniformity in expression of polymer names is a major challenge for widespread use of polymer related data resources and limits broad application of materials informatics for innovation in broad classes of polymer science and polymeric based materials. The current solution of using a variety of different chemical identifiers has proven insufficient to address the challenge and is not intuitive for researchers. This work proposes a multi-algorithm-based mapping methodology entitled ChemProps that is optimized to solve the polymer indexing issue with easy-to-update design both in depth and in width. RESTful API is enabled for lightweight data exchange and easy integration across data systems. A weight factor is assigned to each algorithm to generate scores for candidate chemical names and optimized to maximize the minimum value of the score difference between the ground truth chemical name and the other candidate chemical names. Ten-fold validation is utilized on the 160 training data points to prevent overfitting issues. The obtained set of weight factors achieves a 100% test accuracy on the 54 test data points. The weight factors will evolve as ChemProps grows. With ChemProps, other polymer databases can remove duplicate entries and enable a more accurate “search by SMILES” function by using ChemProps as a common name-to-SMILES translator through API calls. ChemProps is also an excellent tool for auto-populating polymer properties thanks to its easy-to-update design.

scholarly journals USING SEMANTICALLY PAIRED IMAGES TO IMPROVE DOMAIN ADAPTATION FOR THE SEMANTIC SEGMENTATION OF AERIAL IMAGES

2020 ◽  
Vol V-2-2020 ◽  
pp. 483-492
Author(s):  
D. Gritzner ◽  
J. Ostermann
Keyword(s):  
Time Window ◽  
Domain Adaptation ◽  
Geographical Area ◽  
Model Performance ◽  
Ground Truth ◽  
Semantic Segmentation ◽  
Training Data ◽  
Aerial Images ◽  
Target Domain ◽  
Training Examples

Abstract. Modern machine learning, especially deep learning, which is used in a variety of applications, requires a lot of labelled data for model training. Having an insufficient amount of training examples leads to models which do not generalize well to new input instances. This is a particular significant problem for tasks involving aerial images: often training data is only available for a limited geographical area and a narrow time window, thus leading to models which perform poorly in different regions, at different times of day, or during different seasons. Domain adaptation can mitigate this issue by using labelled source domain training examples and unlabeled target domain images to train a model which performs well on both domains. Modern adversarial domain adaptation approaches use unpaired data. We propose using pairs of semantically similar images, i.e., whose segmentations are accurate predictions of each other, for improved model performance. In this paper we show that, as an upper limit based on ground truth, using semantically paired aerial images during training almost always increases model performance with an average improvement of 4.2% accuracy and .036 mean intersection-over-union (mIoU). Using a practical estimate of semantic similarity, we still achieve improvements in more than half of all cases, with average improvements of 2.5% accuracy and .017 mIoU in those cases.

scholarly journals On the objectivity, reliability, and validity of deep learning enabled bioimage analyses

eLife ◽  
2020 ◽  
Vol 9 ◽  
Author(s):  
Dennis Segebarth ◽  
Matthias Griebel ◽  
Nikolai Stein ◽  
Cora R von Collenberg ◽  
Corinna Martin ◽  
...  
Keyword(s):  
Deep Learning ◽  
Signal To Noise Ratio ◽  
Biological Effects ◽  
Reliability And Validity ◽  
Ground Truth ◽  
Training Data ◽  
Model Organisms ◽  
Data Annotation ◽  
Bioimage Analysis ◽  
Model Training

Bioimage analysis of fluorescent labels is widely used in the life sciences. Recent advances in deep learning (DL) allow automating time-consuming manual image analysis processes based on annotated training data. However, manual annotation of fluorescent features with a low signal-to-noise ratio is somewhat subjective. Training DL models on subjective annotations may be instable or yield biased models. In turn, these models may be unable to reliably detect biological effects. An analysis pipeline integrating data annotation, ground truth estimation, and model training can mitigate this risk. To evaluate this integrated process, we compared different DL-based analysis approaches. With data from two model organisms (mice, zebrafish) and five laboratories, we show that ground truth estimation from multiple human annotators helps to establish objectivity in fluorescent feature annotations. Furthermore, ensembles of multiple models trained on the estimated ground truth establish reliability and validity. Our research provides guidelines for reproducible DL-based bioimage analyses.

Glean

2021 ◽  
Vol 14 (6) ◽  
pp. 997-1005
Author(s):  
Sandeep Tata ◽  
Navneet Potti ◽  
James B. Wendt ◽  
Lauro Beltrão Costa ◽  
Marc Najork ◽  
...  
Keyword(s):  
Machine Learning ◽  
Data Management ◽  
Real World ◽  
Empirical Studies ◽  
Ground Truth ◽  
Training Data ◽  
Ground Truth Data ◽  
Document Type ◽  
Machine Learning Model ◽  
Structured Information

Extracting structured information from templatic documents is an important problem with the potential to automate many real-world business workflows such as payment, procurement, and payroll. The core challenge is that such documents can be laid out in virtually infinitely different ways. A good solution to this problem is one that generalizes well not only to known templates such as invoices from a known vendor, but also to unseen ones. We developed a system called Glean to tackle this problem. Given a target schema for a document type and some labeled documents of that type, Glean uses machine learning to automatically extract structured information from other documents of that type. In this paper, we describe the overall architecture of Glean, and discuss three key data management challenges : 1) managing the quality of ground truth data, 2) generating training data for the machine learning model using labeled documents, and 3) building tools that help a developer rapidly build and improve a model for a given document type. Through empirical studies on a real-world dataset, we show that these data management techniques allow us to train a model that is over 5 F1 points better than the exact same model architecture without the techniques we describe. We argue that for such information-extraction problems, designing abstractions that carefully manage the training data is at least as important as choosing a good model architecture.

Sign in / Sign up

Export Citation Format

Share Document