Leveraging Equivariant Features for Absolute Pose Regression

in IEEE Conference on Computer Vision and Pattern Recognition. (2022)

Pose estimation enables vision-based systems to refer to their environment, supporting activities ranging from scene navigation to object manipulation. However, end-to-end approaches, that have achieved ... [more ▼]

Pose estimation enables vision-based systems to refer to their environment, supporting activities ranging from scene navigation to object manipulation. However, end-to-end approaches, that have achieved state-of-the-art performance in many perception tasks, are still unable to compete with 3D geometry-based methods in pose estimation. Indeed, absolute pose regression has been proven to be more related to image retrieval than to 3D structure. Our assumption is that statistical features learned by classical convolutional neural networks do not carry enough geometrical information for reliably solving this task. This paper studies the use of deep equivariant features for end-to-end pose regression. We further propose a translation and rotation equivariant Convolutional Neural Network whose architecture directly induces representations of camera motions into the feature space. In the context of absolute pose regression, this geometric property allows for implicitly augmenting the training data under a whole group of image plane-preserving transformations. Therefore, directly learning equivariant features efficiently compensates for learning intermediate representations that are indirectly equivariant yet data-intensive. Extensive experimental validation demonstrates that our lightweight model outperforms existing ones on standard datasets. [less ▲]

Training very deep neural networks: Rethinking the role of skip connections

in Neurocomputing (2021)

SPACECRAFT RECOGNITION LEVERAGING KNOWLEDGE OF SPACE ENVIRONMENT: SIMULATOR, DATASET, COMPETITION DESIGN, AND ANALYSIS

in 2021 IEEE International Conference on Image Processing (ICIP) (2021)

Towards Generalization of 3D Human Pose Estimation In The Wild

in International Conference on Pattern Recognition (ICPR) Workshop on 3D Human Understanding, Milan 10-15 January 2021 (2020)

In this paper, we propose 3DBodyTex.Pose, a dataset that addresses the task of 3D human pose estimation in-the-wild. Generalization to in-the-wild images remains limited due to the lack of adequate ... [more ▼]

In this paper, we propose 3DBodyTex.Pose, a dataset that addresses the task of 3D human pose estimation in-the-wild. Generalization to in-the-wild images remains limited due to the lack of adequate datasets. Existent ones are usually collected in indoor controlled environments where motion capture systems are used to obtain the 3D ground-truth annotations of humans. 3DBodyTex.Pose offers high quality and rich data containing 405 different real subjects in various clothing and poses, and 81k image samples with ground-truth 2D and 3D pose annotations. These images are generated from 200 viewpoints among which 70 challenging extreme viewpoints. This data was created starting from high resolution textured 3D body scans and by incorporating various realistic backgrounds. Retraining a state-of-the-art 3D pose estimation approach using data augmented with 3DBodyTex.Pose showed promising improvement in the overall performance, and a sensible decrease in the per joint position error when testing on challenging viewpoints. The 3DBodyTex.Pose is expected to offer the research community with new possibilities for generalizing 3D pose estimation from monocular in-the-wild images. [less ▲]

Super-Resolution Approaches for Depth Video Enhancement

Doctoral thesis (2015)

Sensing using 3D technologies has seen a revolution in the past years where cost-effective depth sensors are today part of accessible consumer electronics. Their ability in directly capturing depth videos ... [more ▼]

Sensing using 3D technologies has seen a revolution in the past years where cost-effective depth sensors are today part of accessible consumer electronics. Their ability in directly capturing depth videos in real-time has opened tremendous possibilities for multiple applications in computer vision. These sensors, however, have major shortcomings due to their high noise contamination, including missing and jagged measurements, and their low spatial resolutions. In order to extract detailed 3D features from this type of data, a dedicated data enhancement is required. We propose a generic depth multi-frame super-resolution framework that addresses the limitations of state-of-the-art depth enhancement approaches. The proposed framework does not need any additional hardware or coupling with di erent modalities. It is based on a new data model that uses densely upsampled low resolution observations. This results in a robust median initial estimation, further refined by a deblurring operation using a bilateral total variation as the regularization term. The upsampling operation ensures a systematic improvement in the registration accuracy. This is explored in different scenarios based on the motions involved in the depth video. For the general and most challenging case of objects deforming non-rigidly in full 3D, we propose a recursive dynamic multi-frame super-resolution algorithm where the relative local 3D motions between consecutive frames are directly accounted for. We rely on the assumption that these 3D motions can be decoupled into lateral motions and radial displacements. This allows to perform a simple local per-pixel tracking where both depth measurements and deformations are optimized. As compared to alternative approaches, the results show a clear improvement in reconstruction accuracy and in robustness to noise, to relative large non-rigid deformations, and to topological changes. Moreover, the proposed approach, implemented on a CPU, is shown to be computationally efficient and working in real-time. [less ▲]

Patch-based Statistical Performance Analysis of Upsampling for Precise Super–Resolution

in 11th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISAPP'15) (2015, March)

All existent methods for the statistical analysis of super–resolution approaches have stopped at the variance term, not accounting for the bias in the mean square error. In this paper we give an original ... [more ▼]

All existent methods for the statistical analysis of super–resolution approaches have stopped at the variance term, not accounting for the bias in the mean square error. In this paper we give an original derivation of the bias term. We propose to use a patch-based method inspired by the work of (Chatterjee and Milanfar, 2009). Our approach, however, is completely new as we derive a new affine bias model dedicated for the multi-frame super resolution framework. We apply the proposed statistical performance analysis to the Upsampling for Precise Super–Resolution (UP-SR) algorithm. This algorithm was shown experimentally to be a good solution for enhancing the resolution of depth sequences in both cases of global and local motions. Its performance is herein analyzed theoretically in terms of its approximated mean square error, using the proposed derivation of the bias. This analysis is validated experimentally on simulated static and dynamic depth sequences with a known ground truth. This provides an insightful understanding of the effects of noise variance, number of observed low resolution frames, and super–resolution factor on the final and intermediate performance of UP–SR. Our conclusion is that increasing the number of frames should improve the performance while the error is increased due to local motions, and to the upsampling which is part of UP-SR. [less ▲]

KinectDeform: Enhanced 3D Reconstruction of Non-Rigidly Deforming Objects

in Second International Conference on 3D Vision (2014)

In this work we propose KinectDeform, an algorithm which targets enhanced 3D reconstruction of scenes containing non-rigidly deforming objects. It provides an innovation to the existing class of ... [more ▼]

In this work we propose KinectDeform, an algorithm which targets enhanced 3D reconstruction of scenes containing non-rigidly deforming objects. It provides an innovation to the existing class of algorithms which either target scenes with rigid objects only or allow for very limited non-rigid deformations or use pre-computed templates to track them. KinectDeform combines a fast non-rigid scene tracking algorithm based on octree data representation and hierarchical voxel associations with a recursive data filtering mechanism. We analyze its performance on both real and simulated data and show improved results in terms of smoothness and feature preserving 3D reconstructions with reduced noise. [less ▲]

Multi-frame super-resolution by enhanced shift & add

in International Symposium on Image and Signal Processing and Analysis, ISPA (2013)

A critical step in multi-frame super-resolution is the registration of frames based on their motion. We improve the performance of current state-of-the-art super-resolution techniques by proposing a more ... [more ▼]

A critical step in multi-frame super-resolution is the registration of frames based on their motion. We improve the performance of current state-of-the-art super-resolution techniques by proposing a more robust and accurate registration as early as in the initialization stage of the high resolution estimate. Indeed, we solve the limitations on scale and motion inherent to the classical Shift & Add approach by upsampling the low resolution frames up to the super-resolution factor prior to estimating motion or to median filtering. This is followed by an appropriate selective optimization, leading to an enhanced Shift & Add. Quantitative and qualitative evaluations have been conducted at two levels; the initial estimation and the final optimized superresolution. Results show that the proposed algorithm outperforms existing state-of-art methods. © 2013 University of Trieste and University of Zagreb. [less ▲]

Mutli-Frame Super-Resolution by Enhanced Shift & Add

in 8th International Symposium on Image and Signal Processing and Analysis (2013)

A critical step in multi-frame super-resolution is the registration of frames based on their motion. We improve the performance of current state-of-the-art super-resolution techniques by proposing a more ... [more ▼]

A critical step in multi-frame super-resolution is the registration of frames based on their motion. We improve the performance of current state-of-the-art super-resolution techniques by proposing a more robust and accurate registration as early as in the initialization stage of the high resolution estimate. Indeed, we solve the limitations on scale and motion inherent to the classical Shift & Add approach by upsampling the low resolution frames up to the super-resolution factor prior to estimating motion or to median filtering. This is followed by an appropriate selective optimization, leading to an enhanced Shift & Add. Quantitative and qualitative evaluations have been conducted at two levels; the initial estimation and the final optimized super-resolution. Results show that the proposed algorithm outperforms existing state-of-art methods. [less ▲]

A New Set of Quartic Trivariate Polynomial Equations for Stratied Camera Self-calibrationunder Zero-Skew and Constant Parameters Assumptions

in Computer Vision – ECCV 2012, 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI (2012)

This paper deals with the problem of self-calibrating a moving camera with constant parameters. We propose a new set of quartic trivariate polynomial equations in the unknown coordinates of the plane at ... [more ▼]

This paper deals with the problem of self-calibrating a moving camera with constant parameters. We propose a new set of quartic trivariate polynomial equations in the unknown coordinates of the plane at infinity derived under the no-skew assumption. Our new equations allow to further enforce the constancy of the principal point across all images while retrieving the plane at infinity. Six such polynomials, four of which are independent, are obtained for each triplet of images. The proposed equations can be solved along with the so-called modulus constraints and allow to improve the performance of existing methods. [less ▲]