Andrew I. Comport
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Andrew I. Comport has a total of 25 co-authors. Here are the top co-authors by number of shared publications:
- Adrian Voicila
Renault Software Factory 2600 Route des Crêtes 06560 Valbonne France
2 shared publications - Arnab Dey
I3S-CNRS Université Côte d'Azur
2 shared publications - Jean Martinet
Université Côte d’Azur
2 shared publications - Adrian Voicila
Renault Software Factory, 2600 Rte des Crtes, Valbonne, France
1 shared publication - Andrea Ancora
Renault Group - Ampere, France
1 shared publication - Antitza Dantcheva
Inria Center at Université Côte d'Azur
1 shared publication
Latest Publications
- 16 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Artificial Intelligence
- Vol. 7 (6) , 3574-3584
HFGaussian: Learning Generalizable Gaussian Human With Integrated Human Features
Recent advancements in radiance field rendering show promising results in 3D scene representation, where Gaussian splatting-based techniques emerge as state-of-the-art due to their quality and efficiency. Gaussian splatting is widely used for various applications, including 3D human representation. However, previous 3D Gaussian splatting methods can not render RGB and human features simultaneously. Furthermore, these methods either use parametric body models as additional information or fail to provide any underlying structure, like human biomechanical features, which are essential for various applications such as medical diagnosis. In this paper, we present a novel approach called HFGaussian that can estimate novel views and human features, such as the 3D skeleton, 3D key points, and dense pose, from sparse input images simultaneously in real time at 25 FPS. The proposed method leverages the generalizable Gaussian splatting technique to represent the human subject and 3D consistent human features, enabling efficient and generalizable reconstruction. By incorporating a pose regression network and the feature splatting technique with Gaussian splatting, HFGaussian demonstrates improved capabilities over existing 3D human methods, showcasing the potential of 3D human representations with integrated biomechanics. We thoroughly evaluate our HFGaussian method against the latest state-of-the-art techniques in human Gaussian splatting methods, demonstrating its real-time, state-of-the-art performance. Project page: https://github.com/arnabdeypolimi/HFGaussian_official.
- 16 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Artificial Intelligence
- Vol. 7 (6) , 3574-3584
HFGaussian: Learning Generalizable Gaussian Human With Integrated Human Features
Recent advancements in radiance field rendering show promising results in 3D scene representation, where Gaussian splatting-based techniques emerge as state-of-the-art due to their quality and efficiency. Gaussian splatting is widely used for various applications, including 3D human representation. However, previous 3D Gaussian splatting methods can not render RGB and human features simultaneously. Furthermore, these methods either use parametric body models as additional information or fail to provide any underlying structure, like human biomechanical features, which are essential for various applications such as medical diagnosis. In this paper, we present a novel approach called HFGaussian that can estimate novel views and human features, such as the 3D skeleton, 3D key points, and dense pose, from sparse input images simultaneously in real time at 25 FPS. The proposed method leverages the generalizable Gaussian splatting technique to represent the human subject and 3D consistent human features, enabling efficient and generalizable reconstruction. By incorporating a pose regression network and the feature splatting technique with Gaussian splatting, HFGaussian demonstrates improved capabilities over existing 3D human methods, showcasing the potential of 3D human representations with integrated biomechanics. We thoroughly evaluate our HFGaussian method against the latest state-of-the-art techniques in human Gaussian splatting methods, demonstrating its real-time, state-of-the-art performance. Project page: https://github.com/arnabdeypolimi/HFGaussian_official.
- 5 March 2025
- Published by American Physical Society (APS) in Physical Review E
- Vol. 111 (3) , 035301
AI-based computation method for the Eddington factor in the M1-multigroup model
Radiative hydrodynamics models the interaction between fluid flows and radiation, which is essential for many astrophysical simulations. The M1-multigroup model is widely regarded as the most precise framework for capturing the complex interplay between light and matter, particularly by accounting for the spectral behavior of photons. A critical component of this model is the Eddington factor, which is used in the closure relation linking the radiative pressure to the radiative energy. Although an analytical expression for the Eddington factor does not exist, our research reveals that it depends solely on three parameters: the radiative temperature, the reduced flux, and the group narrowness. To address the challenge of calculating this factor efficiently, we have developed a method that combines neural networks and polynomial approximations. This method achieves computational speeds up to 3000 times faster than traditional line search algorithms while providing precision levels up to 1000 times higher than simplified alternatives based on interpolation or the analytical expression of the M1-gray model. Unlike interpolation-based techniques, it operates without requiring prior knowledge of radiative quantities, offering greater flexibility and applicability in diverse scenarios. Although the test simulations we performed, where radiation pressure is not the dominant factor, demonstrate a limited impact of the precision of the estimation of the Eddington factor on fluid dynamics, our approach lays a solid foundation for future advancements. It is particularly promising for more complex simulations involving out-of-equilibrium, radiative pressure-dominated scenarios. These developments mark a significant step forward in radiative hydrodynamics, offering a robust, accurate, and highly efficient computational tool to advance astrophysical modeling.
- 5 March 2025
- Published by American Physical Society (APS) in Physical Review E
- Vol. 111 (3) , 035301
AI-based computation method for the Eddington factor in the M1-multigroup model
Radiative hydrodynamics models the interaction between fluid flows and radiation, which is essential for many astrophysical simulations. The M1-multigroup model is widely regarded as the most precise framework for capturing the complex interplay between light and matter, particularly by accounting for the spectral behavior of photons. A critical component of this model is the Eddington factor, which is used in the closure relation linking the radiative pressure to the radiative energy. Although an analytical expression for the Eddington factor does not exist, our research reveals that it depends solely on three parameters: the radiative temperature, the reduced flux, and the group narrowness. To address the challenge of calculating this factor efficiently, we have developed a method that combines neural networks and polynomial approximations. This method achieves computational speeds up to 3000 times faster than traditional line search algorithms while providing precision levels up to 1000 times higher than simplified alternatives based on interpolation or the analytical expression of the M1-gray model. Unlike interpolation-based techniques, it operates without requiring prior knowledge of radiative quantities, offering greater flexibility and applicability in diverse scenarios. Although the test simulations we performed, where radiation pressure is not the dominant factor, demonstrate a limited impact of the precision of the estimation of the Eddington factor on fluid dynamics, our approach lays a solid foundation for future advancements. It is particularly promising for more complex simulations involving out-of-equilibrium, radiative pressure-dominated scenarios. These developments mark a significant step forward in radiative hydrodynamics, offering a robust, accurate, and highly efficient computational tool to advance astrophysical modeling.
- Open Access
- 24 May 2024
- Published by Association for Computing Machinery (ACM)
- p. 204-208
Low Parameter Neural Networks for In-Car Distracted Driver Detection
In response to the escalating global trend of increasing road accidents, particularly those attributed to distracted driving, this study endeavors to construct a precise and robust system for detecting distracted drivers and consequently issuing timely warnings. We propose a CNN-based framework designed not only to identify distracted drivers but also to determine the specific cause of the distraction. Experimental findings reveal that our system surpasses previous methodologies in the literature, achieving an accuracy of 99.31% with the ensemble method and 99.46% with our Convolution Neural Network. To ensure the practical applicability and adoption of our models within the car manufacturing industry, we designed them with a focus on minimizing the number of parameters. Furthermore, we present our architectures with a substantial reduction in the number of parameters from 140 million in the original VGG-16 to approximately 200 thousand.
- 24 May 2024
- Published by Association for Computing Machinery (ACM)
- p. 204-208
Open AccessLow Parameter Neural Networks for In-Car Distracted Driver Detection
In response to the escalating global trend of increasing road accidents, particularly those attributed to distracted driving, this study endeavors to construct a precise and robust system for detecting distracted drivers and consequently issuing timely warnings. We propose a CNN-based framework designed not only to identify distracted drivers but also to determine the specific cause of the distraction. Experimental findings reveal that our system surpasses previous methodologies in the literature, achieving an accuracy of 99.31% with the ensemble method and 99.46% with our Convolution Neural Network. To ensure the practical applicability and adoption of our models within the car manufacturing industry, we designed them with a focus on minimizing the number of parameters. Furthermore, we present our architectures with a substantial reduction in the number of parameters from 140 million in the original VGG-16 to approximately 200 thousand.
- Open Access
GHNeRF: Learning Generalizable Human Features with Efficient Neural Radiance Fields
Recent advances in Neural Radiance Fields (NeRF) have demonstrated promising results in 3D scene representations, including 3D human representations. However, these representations often lack crucial information on the underlying human pose and structure, which is crucial for AR/VR applications and games. In this paper, we introduce a novel approach, termed GHNeRF, designed to address these limitations by learning 2D/3D joint locations of human subjects with NeRF representation. GHNeRF uses a pre-trained 2D encoder streamlined to extract essential human features from 2D images, which are then incorporated into the NeRF framework in order to encode human biomechanic features. This allows our network to simultaneously learn biomechanic features, such as joint locations, along with human geometry and texture. To assess the effectiveness of our method, we conduct a comprehensive comparison with state-of-the-art human NeRF techniques and joint estimation algorithms. Our results show that GHNeRF can achieve state-of-the-art results in near real-time.
Open AccessGHNeRF: Learning Generalizable Human Features with Efficient Neural Radiance Fields
Recent advances in Neural Radiance Fields (NeRF) have demonstrated promising results in 3D scene representations, including 3D human representations. However, these representations often lack crucial information on the underlying human pose and structure, which is crucial for AR/VR applications and games. In this paper, we introduce a novel approach, termed GHNeRF, designed to address these limitations by learning 2D/3D joint locations of human subjects with NeRF representation. GHNeRF uses a pre-trained 2D encoder streamlined to extract essential human features from 2D images, which are then incorporated into the NeRF framework in order to encode human biomechanic features. This allows our network to simultaneously learn biomechanic features, such as joint locations, along with human geometry and texture. To assess the effectiveness of our method, we conduct a comprehensive comparison with state-of-the-art human NeRF techniques and joint estimation algorithms. Our results show that GHNeRF can achieve state-of-the-art results in near real-time.
- 4 April 2024
- Published by Elsevier in Ocean Engineering
- Vol. 302
MAMSLAM: Towards underwater-robust multi-agent visual SLAM- 4 April 2024
- Published by Elsevier in Ocean Engineering
- Vol. 302
MAMSLAM: Towards underwater-robust multi-agent visual SLAM- Open Access
RGB-D Neural Radiance Fields: Local Sampling for Faster Training
Learning a 3D representation of a scene has been a challenging problem for decades in computer vision. Recent advances in implicit neural representation from images using neural radiance fields(NeRF) have shown promising results. Some of the limitations of previous NeRF based methods include longer training time, and inaccurate underlying geometry. The proposed method takes advantage of RGB-D data to reduce training time by leveraging depth sensing to improve local sampling. This paper proposes a depth-guided local sampling strategy and a smaller neural network architecture to achieve faster training time without compromising quality.
Open AccessRGB-D Neural Radiance Fields: Local Sampling for Faster Training
Learning a 3D representation of a scene has been a challenging problem for decades in computer vision. Recent advances in implicit neural representation from images using neural radiance fields(NeRF) have shown promising results. Some of the limitations of previous NeRF based methods include longer training time, and inaccurate underlying geometry. The proposed method takes advantage of RGB-D data to reduce training time by leveraging depth sensing to improve local sampling. This paper proposes a depth-guided local sampling strategy and a smaller neural network architecture to achieve faster training time without compromising quality.
- Open Access
Instance-aware multi-object self-supervision for monocular depth prediction
This paper proposes a self-supervised monocular image-to-depth prediction framework that is trained with an end-to-end photometric loss that handles not only 6-DOF camera motion but also 6-DOF moving object instances. Self-supervision is performed by warping the images across a video sequence using depth and scene motion including object instances. One novelty of the proposed method is the use of the multi-head attention of the transformer network that matches moving objects across time and models their interaction and dynamics. This enables accurate and robust pose estimation for each object instance. Most image-to-depth predication frameworks make the assumption of rigid scenes, which largely degrades their performance with respect to dynamic objects. Only a few SOTA papers have accounted for dynamic objects. The proposed method is shown to outperform these methods on standard benchmarks and the impact of the dynamic motion on these benchmarks is exposed. Furthermore, the proposed image-to-depth prediction framework is also shown to be competitive with SOTA video-to-depth prediction frameworks.
Open AccessInstance-aware multi-object self-supervision for monocular depth prediction
This paper proposes a self-supervised monocular image-to-depth prediction framework that is trained with an end-to-end photometric loss that handles not only 6-DOF camera motion but also 6-DOF moving object instances. Self-supervision is performed by warping the images across a video sequence using depth and scene motion including object instances. One novelty of the proposed method is the use of the multi-head attention of the transformer network that matches moving objects across time and models their interaction and dynamics. This enables accurate and robust pose estimation for each object instance. Most image-to-depth predication frameworks make the assumption of rigid scenes, which largely degrades their performance with respect to dynamic objects. Only a few SOTA papers have accounted for dynamic objects. The proposed method is shown to outperform these methods on standard benchmarks and the impact of the dynamic motion on these benchmarks is exposed. Furthermore, the proposed image-to-depth prediction framework is also shown to be competitive with SOTA video-to-depth prediction frameworks.
- Open Access
Forecasting of depth and ego-motion with transformers and self-supervision
This paper addresses the problem of end-to-end self-supervised forecasting of depth and ego motion. Given a sequence of raw images, the aim is to forecast both the geometry and ego-motion using a self supervised photometric loss. The architecture is designed using both convolution and transformer modules. This leverages the benefits of both modules: Inductive bias of CNN, and the multi-head attention of transformers, thus enabling a rich spatio-temporal representation that enables accurate depth forecasting. Prior work attempts to solve this problem using multi-modal input/output with supervised ground-truth data which is not practical since a large annotated dataset is required. Alternatively to prior methods, this paper forecasts depth and ego motion using only self-supervised raw images as input. The approach performs significantly well on the KITTI dataset benchmark with several performance criteria being even comparable to prior non-forecasting self-supervised monocular depth inference methods.
Open AccessForecasting of depth and ego-motion with transformers and self-supervision
This paper addresses the problem of end-to-end self-supervised forecasting of depth and ego motion. Given a sequence of raw images, the aim is to forecast both the geometry and ego-motion using a self supervised photometric loss. The architecture is designed using both convolution and transformer modules. This leverages the benefits of both modules: Inductive bias of CNN, and the multi-head attention of transformers, thus enabling a rich spatio-temporal representation that enables accurate depth forecasting. Prior work attempts to solve this problem using multi-modal input/output with supervised ground-truth data which is not practical since a large annotated dataset is required. Alternatively to prior methods, this paper forecasts depth and ego motion using only self-supervised raw images as input. The approach performs significantly well on the KITTI dataset benchmark with several performance criteria being even comparable to prior non-forecasting self-supervised monocular depth inference methods.
- Open Access
Are conditional GANs explicitly conditional?
This paper proposes two important contributions for conditional Generative Adversarial Networks (cGANs) to improve the wide variety of applications that exploit this architecture. The first main contribution is an analysis of cGANs to show that they are not explicitly conditional. In particular, it will be shown that the discriminator and subsequently the cGAN does not automatically learn the conditionality between inputs. The second contribution is a new method, called a contrario cGAN, that explicitly models conditionality for both parts of the adversarial architecture via a novel a contrario loss that involves training the discriminator to learn unconditional (adverse) examples. This leads to a novel type of data augmentation approach for GANs (a contrario learning) which allows to restrict the search space of the generator to conditional outputs using adverse examples. Extensive experimentation is carried out to evaluate the conditionality of the discriminator by proposing a probability distribution analysis. Comparisons with the cGAN architecture for different applications show significant improvements in performance on well known datasets including, semantic image synthesis, image segmentation, monocular depth prediction and "single label"-to-image using different metrics including Fr\'echet Inception Distance (FID), mean Intersection over Union (mIoU), Root Mean Square Error log (RMSE log) and Number of statistically-Different Bins (NDB).
Open AccessAre conditional GANs explicitly conditional?
This paper proposes two important contributions for conditional Generative Adversarial Networks (cGANs) to improve the wide variety of applications that exploit this architecture. The first main contribution is an analysis of cGANs to show that they are not explicitly conditional. In particular, it will be shown that the discriminator and subsequently the cGAN does not automatically learn the conditionality between inputs. The second contribution is a new method, called a contrario cGAN, that explicitly models conditionality for both parts of the adversarial architecture via a novel a contrario loss that involves training the discriminator to learn unconditional (adverse) examples. This leads to a novel type of data augmentation approach for GANs (a contrario learning) which allows to restrict the search space of the generator to conditional outputs using adverse examples. Extensive experimentation is carried out to evaluate the conditionality of the discriminator by proposing a probability distribution analysis. Comparisons with the cGAN architecture for different applications show significant improvements in performance on well known datasets including, semantic image synthesis, image segmentation, monocular depth prediction and "single label"-to-image using different metrics including Fr\'echet Inception Distance (FID), mean Intersection over Union (mIoU), Root Mean Square Error log (RMSE log) and Number of statistically-Different Bins (NDB).
- 15 September 2014
- Published by Wiley in Journal of Field Robotics
- Vol. 32 (4) , 474-503
Dense Omnidirectional RGB‐D Mapping of Large‐scale Outdoor Environments for Real‐time Localization and Autonomous Navigation
This paper presents a novel method and innovative apparatus for building three‐dimensional (3D) dense visual maps of large‐scale unstructured environments for autonomous navigation and real‐time localization. The main contribution of the paper is focused on proposing an efficient and accurate 3D world representation that allows us to extend the boundaries of state‐of‐the‐art dense visual mapping to large scales. This is achieved via an omnidirectional key‐frame representation of the environment, which is able to synthesize photorealistic views of captured environments at arbitrary locations. Locally, the representation is image‐based (egocentric) and is composed of accurate augmented spherical panoramas combining photometric information (RGB), depth information (D), and saliency for all viewing directions at a particular point in space (i.e., a point in the light field). The spheres are related by a graph of six degree of freedom (DOF) poses (3 DOF translation and 3 DOF rotation) that are estimated through multiview spherical registration. It is shown that this world representation can be used to perform robust real‐time localization (in 6 DOF) of any configuration of visual sensors within their environment, whether they be monocular, stereo, or multiview. Contrary to feature‐based approaches, an efficient direct image registration technique is formulated. This approach directly exploits the advantages of the spherical representation by minimizing a photometric error between a current image and a reference sphere. Two novel multicamera acquisition systems have been developed and calibrated to acquire this information, and this paper reports for the first time the second system. Given the robustness and efficiency of this representation, field experiments demonstrating autonomous navigation and large‐scale mapping will be reported in detail for challenging unstructured environments, containing vegetation, pedestrians, varying illumination conditions, trams, and dense traffic.
- 15 September 2014
- Published by Wiley in Journal of Field Robotics
- Vol. 32 (4) , 474-503
Dense Omnidirectional RGB‐D Mapping of Large‐scale Outdoor Environments for Real‐time Localization and Autonomous Navigation
This paper presents a novel method and innovative apparatus for building three‐dimensional (3D) dense visual maps of large‐scale unstructured environments for autonomous navigation and real‐time localization. The main contribution of the paper is focused on proposing an efficient and accurate 3D world representation that allows us to extend the boundaries of state‐of‐the‐art dense visual mapping to large scales. This is achieved via an omnidirectional key‐frame representation of the environment, which is able to synthesize photorealistic views of captured environments at arbitrary locations. Locally, the representation is image‐based (egocentric) and is composed of accurate augmented spherical panoramas combining photometric information (RGB), depth information (D), and saliency for all viewing directions at a particular point in space (i.e., a point in the light field). The spheres are related by a graph of six degree of freedom (DOF) poses (3 DOF translation and 3 DOF rotation) that are estimated through multiview spherical registration. It is shown that this world representation can be used to perform robust real‐time localization (in 6 DOF) of any configuration of visual sensors within their environment, whether they be monocular, stereo, or multiview. Contrary to feature‐based approaches, an efficient direct image registration technique is formulated. This approach directly exploits the advantages of the spherical representation by minimizing a photometric error between a current image and a reference sphere. Two novel multicamera acquisition systems have been developed and calibrated to acquire this information, and this paper reports for the first time the second system. Given the robustness and efficiency of this representation, field experiments demonstrating autonomous navigation and large‐scale mapping will be reported in detail for challenging unstructured environments, containing vegetation, pedestrians, varying illumination conditions, trams, and dense traffic.