Sagi Filin
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Sagi Filin has a total of 36 co-authors. Here are the top co-authors by number of shared publications:
- Bashar Elnashef
Mapping and Geo-Information Engineering, Technion – Israel Institute of Technology, Haifa, Israel
11 shared publications - Reuma Arav
Institute of Geomatics, Dept. of Ecosystem Management, Climate, and Biodiversity, BOKU University, 1190, Vienna, Austria
6 shared publications - Yoav Avni
Geological Survey of Israel
3 shared publications - F. Xue
Mapping and Geoinformation Engineering, Technion –Israel Institute of Technology, Haifa, Israel
2 shared publications Grigory Ilizirov
Technion City, Mapping & Geoinformat Engn, IL-32000 Haifa, Israel
2 shared publications- Itzik Klein
University of Haifa
2 shared publications
Latest Publications
- Open Access
- 3 November 2025
- Published by Copernicus GmbH in ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
- Vol. X-1/W2-202, 19-25
Enhancing Rotation Averaging and Global Positioning by an Adaptive Robust Kernel
Motion averaging (MA) offers an efficient and mostly linear means for estimating image sets pose and provides a reliable initialization for large-scale structure-from-motion (SfM) pipelines. Nonetheless, MA, which comprises rotation and translation averaging (RA & TA, respectively), can be severely affected by the presence of outliers that can greatly degrade the performance of the entire optimization process or even lead to the divergence of the SfM solution. While robust loss functions have been applied to mitigate outliers effect, their performance depends heavily on the choice of parameters and requires manual tuning and prior knowledge of the residual distribution. To make MA more robust, we enhance it by incorporating an adaptive robust kernel that automatically adjusts its parameter to the residual distribution. This adaptive behavior balances robustness and sensitivity, removing the need for manual parameter tuning. In addition, to address the ill-posed nature of TA, we adopt a global positioning framework that jointly estimates the camera and 3D point positions. Experimental results show how the adaptive robust kernel consistently outperforms state-of-the-art fixed-parameter functions. It improves accuracy in both RA and global positioning, particularly in scenes with high levels of noise or outliers. These results demonstrate the effectiveness of adaptive robust kernels for improving the reliability and generalization of MA pipelines in challenging reconstruction scenarios.
- 3 November 2025
- Published by Copernicus GmbH in ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
- Vol. X-1/W2-202, 19-25
Open AccessEnhancing Rotation Averaging and Global Positioning by an Adaptive Robust Kernel
Motion averaging (MA) offers an efficient and mostly linear means for estimating image sets pose and provides a reliable initialization for large-scale structure-from-motion (SfM) pipelines. Nonetheless, MA, which comprises rotation and translation averaging (RA & TA, respectively), can be severely affected by the presence of outliers that can greatly degrade the performance of the entire optimization process or even lead to the divergence of the SfM solution. While robust loss functions have been applied to mitigate outliers effect, their performance depends heavily on the choice of parameters and requires manual tuning and prior knowledge of the residual distribution. To make MA more robust, we enhance it by incorporating an adaptive robust kernel that automatically adjusts its parameter to the residual distribution. This adaptive behavior balances robustness and sensitivity, removing the need for manual parameter tuning. In addition, to address the ill-posed nature of TA, we adopt a global positioning framework that jointly estimates the camera and 3D point positions. Experimental results show how the adaptive robust kernel consistently outperforms state-of-the-art fixed-parameter functions. It improves accuracy in both RA and global positioning, particularly in scenes with high levels of noise or outliers. These results demonstrate the effectiveness of adaptive robust kernels for improving the reliability and generalization of MA pipelines in challenging reconstruction scenarios.
- Open Access
- 4 October 2025
- Published by Copernicus GmbH in The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
- Vol. XLVIII-M-9, 1699-1704
Deep Contour Detection for Enhanced Heritage Sites Visualization by Laser Point Clouds
Laser scanning is a key tool for capturing detailed 3D representations of cultural heritage sites, enabling applications in visualization, conservation, and analysis. However, the resulting point clouds comprise of visually monotonous surfaces with little intuitive value for direct interpretation. Conventional representations—such as collections of surfaces, 3D lines, or boundary-based topological models—often lack the expressiveness or adaptability needed to capture semantically rich, human-intuitive features like salient architectural regions. These salient regions serve as interpretable medium to understand the scan content but remain difficult to extract. Therefore, this paper proposes a framework that learns salient region-aware representations from raw point clouds. Central to the proposed approach is a heat-propagation-driven graph network trained with a new spectral supervision signal. To ensure scalability, our network operates on local patches and introduces a prediction aggregation scheme that efficiently scales to heritage-scale scenes with millions of points. Results show superior performance over existing salient region depiction approaches, producing high-quality results in under two seconds. It is orders of magnitude faster than state-of-the-art methods, even for large-scale noisy terrestrial scans of millions of points.
- 4 October 2025
- Published by Copernicus GmbH in The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
- Vol. XLVIII-M-9, 1699-1704
Open AccessDeep Contour Detection for Enhanced Heritage Sites Visualization by Laser Point Clouds
Laser scanning is a key tool for capturing detailed 3D representations of cultural heritage sites, enabling applications in visualization, conservation, and analysis. However, the resulting point clouds comprise of visually monotonous surfaces with little intuitive value for direct interpretation. Conventional representations—such as collections of surfaces, 3D lines, or boundary-based topological models—often lack the expressiveness or adaptability needed to capture semantically rich, human-intuitive features like salient architectural regions. These salient regions serve as interpretable medium to understand the scan content but remain difficult to extract. Therefore, this paper proposes a framework that learns salient region-aware representations from raw point clouds. Central to the proposed approach is a heat-propagation-driven graph network trained with a new spectral supervision signal. To ensure scalability, our network operates on local patches and introduces a prediction aggregation scheme that efficiently scales to heritage-scale scenes with millions of points. Results show superior performance over existing salient region depiction approaches, producing high-quality results in under two seconds. It is orders of magnitude faster than state-of-the-art methods, even for large-scale noisy terrestrial scans of millions of points.
- Open Access
- 14 July 2025
- Published by Copernicus GmbH in ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
- Vol. X-G-2025, 1037-1044
Edge-aware joint neural denoising and normal estimation for mobile and handheld laser point clouds
Portable laser scanners, handheld and mobile, have gained popularity for their ability to rapidly and economically document scenes. However, the acquired data are characterized by hight levels of noise and by low resolution, both affecting their consequent analysis and 3D modeling. It is customary to enhance their quality by denoising the data by means of point position updates where current approaches independently predict the displacement per point. Such strategies neglect local structural consistency and often yield a non-smooth outcome. To address these shortcomings this paper formulates denoising by local contextual relationships and point assignment to the underlying surface in an end-to-end framework. To extract contextual information it introduces densely packed graph convolution layers and a global attention mechanism. Realizing also that utilization of the conventional L2-norm-driven approaches tends to oversmooth the surface, it introduces a novel bilateral loss that not only mitigates the noise but also preserves sharp geometric features. As a result, our newly developed network learns a shape context representation that measures neighbor similarity and contributes to more accurate surface normals. Performance analysis demonstrates that over 93% of points deviate by ≤ 1 cm – double the percentage achieved by state-of-the-art denoising networks.
- 14 July 2025
- Published by Copernicus GmbH in ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
- Vol. X-G-2025, 1037-1044
Open AccessEdge-aware joint neural denoising and normal estimation for mobile and handheld laser point clouds
Portable laser scanners, handheld and mobile, have gained popularity for their ability to rapidly and economically document scenes. However, the acquired data are characterized by hight levels of noise and by low resolution, both affecting their consequent analysis and 3D modeling. It is customary to enhance their quality by denoising the data by means of point position updates where current approaches independently predict the displacement per point. Such strategies neglect local structural consistency and often yield a non-smooth outcome. To address these shortcomings this paper formulates denoising by local contextual relationships and point assignment to the underlying surface in an end-to-end framework. To extract contextual information it introduces densely packed graph convolution layers and a global attention mechanism. Realizing also that utilization of the conventional L2-norm-driven approaches tends to oversmooth the surface, it introduces a novel bilateral loss that not only mitigates the noise but also preserves sharp geometric features. As a result, our newly developed network learns a shape context representation that measures neighbor similarity and contributes to more accurate surface normals. Performance analysis demonstrates that over 93% of points deviate by ≤ 1 cm – double the percentage achieved by state-of-the-art denoising networks.
- Open Access
- 20 January 2025
- Published by Copernicus GmbH
Using current 3D point clouds as a tool to infer on past geomorphological processes
Examining deposition and erosion dynamics during the late Pleistocene and Holocene is crucial for gaining insights into soil development, erosion, and climate fluctuations. This urgency intensifies as arable lands face escalating degradation rates, particularly in arid and semi-arid environments. Nevertheless, as the destructive nature of erosional processes allows only for short-term studies, long-term processes in these regions are insufficiently investigated. In that respect, the ancient agricultural installations in the arid Southern Levant offer distinctive and undisturbed evidence of long-term land dynamics. Constructed on a late Pleistocene fluvial-loess section during the 3rd-4th CE and abandoned after 600-700 years, these installations record sediment deposition, soil formation, and erosion processes. The challenge is to trace and quantify these processes based on their current state. In this presentation, we demonstrate how the use of 3D point cloud data enables us to follow past geomorphological processes and reconstruct trends and rates. Utilizing data gathered in the immediate vicinity of the UNESCO World Heritage Site of Avdat (Israel), we illustrate how these point clouds comprehensively document the history of soil dynamics in the region. This encompasses the initial erosion phase, subsequent soil aggradation processes resulting from anthropogenic interruption, and the ongoing reinstated erosion. The unique setting, which uncovers the different fluvial sections, together with the detailed 3D documentation of the site, allows us to develop means for the reconstruction of the natural environment in each of the erosion/siltation stages. Therefore, by utilizing the obtained data, we can recreate the site during its developmental stages till the present day. Furthermore, we utilize terrestrial laser scan data sequence acquired in the past decade (2012-2022) to compute current erosion rates. These are then used to determine past rates, enabling inferences about the climatic conditions prevalent in the region over the last millennium. The in-depth examination of these installations provides valuable insights into approaches for soil conservation, sustainable desert living, and strategies to safeguard world-heritage sites subjected to soil erosion. As the global imperative to address soil erosion intensifies, this case study gains heightened relevance.
- 20 January 2025
- Published by Copernicus GmbH
Open AccessUsing current 3D point clouds as a tool to infer on past geomorphological processes
Examining deposition and erosion dynamics during the late Pleistocene and Holocene is crucial for gaining insights into soil development, erosion, and climate fluctuations. This urgency intensifies as arable lands face escalating degradation rates, particularly in arid and semi-arid environments. Nevertheless, as the destructive nature of erosional processes allows only for short-term studies, long-term processes in these regions are insufficiently investigated. In that respect, the ancient agricultural installations in the arid Southern Levant offer distinctive and undisturbed evidence of long-term land dynamics. Constructed on a late Pleistocene fluvial-loess section during the 3rd-4th CE and abandoned after 600-700 years, these installations record sediment deposition, soil formation, and erosion processes. The challenge is to trace and quantify these processes based on their current state. In this presentation, we demonstrate how the use of 3D point cloud data enables us to follow past geomorphological processes and reconstruct trends and rates. Utilizing data gathered in the immediate vicinity of the UNESCO World Heritage Site of Avdat (Israel), we illustrate how these point clouds comprehensively document the history of soil dynamics in the region. This encompasses the initial erosion phase, subsequent soil aggradation processes resulting from anthropogenic interruption, and the ongoing reinstated erosion. The unique setting, which uncovers the different fluvial sections, together with the detailed 3D documentation of the site, allows us to develop means for the reconstruction of the natural environment in each of the erosion/siltation stages. Therefore, by utilizing the obtained data, we can recreate the site during its developmental stages till the present day. Furthermore, we utilize terrestrial laser scan data sequence acquired in the past decade (2012-2022) to compute current erosion rates. These are then used to determine past rates, enabling inferences about the climatic conditions prevalent in the region over the last millennium. The in-depth examination of these installations provides valuable insights into approaches for soil conservation, sustainable desert living, and strategies to safeguard world-heritage sites subjected to soil erosion. As the global imperative to address soil erosion intensifies, this case study gains heightened relevance.
- 18 October 2023
- Published by Elsevier in ISPRS Journal of Photogrammetry and Remote Sensing
- Vol. 205, 263-283
Spatio-temporal registration of plants non-rigid 3-D structure
- 18 October 2023
- Published by Elsevier in ISPRS Journal of Photogrammetry and Remote Sensing
- Vol. 205, 263-283
Spatio-temporal registration of plants non-rigid 3-D structure
- Open Access
- 26 June 2023
- Published by Copernicus GmbH in The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
- Vol. XLVIII-M-2, 1653-1658
SHAPE RECONSTRUCTION OF HERITAGE ASSETS BY POLARIZATION INFORMATION
Three-dimensional modeling of heritage assets poses a challenge when high reflectance and featureless surfaces are involved. Because of the complex reflection characteristics, traditional methods, either passive or active, tend to generate noisy and incomplete reconstruction results. To address this, we introduce in this paper a shape-from-polarization system and modeling strategy for the 3-D reconstruction of heritage assets. As demonstrated, by analyzing polarization properties of reflected light, we can generate by a single image pixel-level-resolution normal and depth maps. We consider two system prototypes involving a point light source but different in the polarizing camera configurations. They provide solutions for different reconstruction requirements and acquisition conditions. Our reconstruction model is common however to both involving a closed set of simple computation steps. Our evaluations demonstrate how the proposed method significantly improves the 3-D surface model completeness and level of detail, showing its suitability and value to reach the goal of a high-resolution survey of heritage assets, where traditional approaches tend to fail.
- 26 June 2023
- Published by Copernicus GmbH in The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
- Vol. XLVIII-M-2, 1653-1658
Open AccessSHAPE RECONSTRUCTION OF HERITAGE ASSETS BY POLARIZATION INFORMATION
Three-dimensional modeling of heritage assets poses a challenge when high reflectance and featureless surfaces are involved. Because of the complex reflection characteristics, traditional methods, either passive or active, tend to generate noisy and incomplete reconstruction results. To address this, we introduce in this paper a shape-from-polarization system and modeling strategy for the 3-D reconstruction of heritage assets. As demonstrated, by analyzing polarization properties of reflected light, we can generate by a single image pixel-level-resolution normal and depth maps. We consider two system prototypes involving a point light source but different in the polarizing camera configurations. They provide solutions for different reconstruction requirements and acquisition conditions. Our reconstruction model is common however to both involving a closed set of simple computation steps. Our evaluations demonstrate how the proposed method significantly improves the 3-D surface model completeness and level of detail, showing its suitability and value to reach the goal of a high-resolution survey of heritage assets, where traditional approaches tend to fail.
- 16 June 2023
- Published by Elsevier in ISPRS Journal of Photogrammetry and Remote Sensing
- Vol. 202, 142-157
Polarimetric monocular leaf normal estimation model for plant phenotyping
- 16 June 2023
- Published by Elsevier in ISPRS Journal of Photogrammetry and Remote Sensing
- Vol. 202, 142-157
Polarimetric monocular leaf normal estimation model for plant phenotyping
- 21 March 2023
- Published by Elsevier in ISPRS Journal of Photogrammetry and Remote Sensing
- Vol. 198, 223-237
A three-point solution with scale estimation ability for two-view flat-refractive underwater photogrammetry
- 21 March 2023
- Published by Elsevier in ISPRS Journal of Photogrammetry and Remote Sensing
- Vol. 198, 223-237
A three-point solution with scale estimation ability for two-view flat-refractive underwater photogrammetry
- 16 March 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Robotics and Automation Letters
- Vol. 8 (5) , 2676-2683
Drift Reduction in Underwater Egomotion Computation by Axial Camera Modeling
Localization of autonomous underwater vehicles is essential for underwater exploration, environmental monitoring, and infrastructure inspections, as only a few examples. As wirelessly transmitted data cannot be used in underwater navigation, and equivalent sensors are costly and not readily deployable, the use of cameras as a localization aid offers an attractive alternative. Nonetheless, the utilization of cameras for underwater navigation is hindered by medium-induced visual deterioration and refraction of the incident ray. While most research has been focusing on addressing the radiometric deterioration effects, we demonstrate that the introduction of a physically aware model of the refraction effect can significantly improve the platform localization. Specifically, we develop a refraction-aware continuous formulation of the egomotion model, which computes the translational and rotational velocities using optical flow measurements. We also demonstrate that a linear model can be reached despite the nonlinear ray's path. Results show an improved navigation solution echoed by a significant reduction in the drift with more than tenfold improvement in the estimated position and angular quantities compared to state-of-the-art methods. They demonstrate the benefit of exact modeling in underwater navigation.
- 16 March 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Robotics and Automation Letters
- Vol. 8 (5) , 2676-2683
Drift Reduction in Underwater Egomotion Computation by Axial Camera Modeling
Localization of autonomous underwater vehicles is essential for underwater exploration, environmental monitoring, and infrastructure inspections, as only a few examples. As wirelessly transmitted data cannot be used in underwater navigation, and equivalent sensors are costly and not readily deployable, the use of cameras as a localization aid offers an attractive alternative. Nonetheless, the utilization of cameras for underwater navigation is hindered by medium-induced visual deterioration and refraction of the incident ray. While most research has been focusing on addressing the radiometric deterioration effects, we demonstrate that the introduction of a physically aware model of the refraction effect can significantly improve the platform localization. Specifically, we develop a refraction-aware continuous formulation of the egomotion model, which computes the translational and rotational velocities using optical flow measurements. We also demonstrate that a linear model can be reached despite the nonlinear ray's path. Results show an improved navigation solution echoed by a significant reduction in the drift with more than tenfold improvement in the estimated position and angular quantities compared to state-of-the-art methods. They demonstrate the benefit of exact modeling in underwater navigation.
- 23 February 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 23 (7) , 7750-7757
Relative Constraints and Their Contribution to Image Configurations
Simultaneous localization of a group of mobile platforms has seen growing interest in recent years. It is common in such setups to equip each platform with a high-end inertial measurement unit and an imaging system. Notwithstanding, present research into such platforms’ dynamics tends to introduce some forms of relaxation, e.g., planar motion, ideal inertial sensors, or a static and near-sight scene. Yet, such assumptions are violated in real-world scenarios, for example, cars driving on bumpy roads, or operating with low-performance inertial sensors. In this paper, we propose to improve the navigation solution of each member in the mobile platforms group and reflect actual operational conditions by the introduction of stochastic constraints on the relative pose of the platforms. Our framework improves the accuracy of the navigation solution and demonstrates greater flexibility in mission–planning and a shorter time on site. Our results show that accurate pose estimates are maintained even when the imaging scale is increased, allowing greater coverage and shorter operation time, which in turn reduces costs. This is demonstrated by an analytical assessment, simulations, and real–world experiments on a group of platforms equipped with low–cost, off–the–shelf sensors.
- 23 February 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 23 (7) , 7750-7757
Relative Constraints and Their Contribution to Image Configurations
Simultaneous localization of a group of mobile platforms has seen growing interest in recent years. It is common in such setups to equip each platform with a high-end inertial measurement unit and an imaging system. Notwithstanding, present research into such platforms’ dynamics tends to introduce some forms of relaxation, e.g., planar motion, ideal inertial sensors, or a static and near-sight scene. Yet, such assumptions are violated in real-world scenarios, for example, cars driving on bumpy roads, or operating with low-performance inertial sensors. In this paper, we propose to improve the navigation solution of each member in the mobile platforms group and reflect actual operational conditions by the introduction of stochastic constraints on the relative pose of the platforms. Our framework improves the accuracy of the navigation solution and demonstrates greater flexibility in mission–planning and a shorter time on site. Our results show that accurate pose estimates are maintained even when the imaging scale is increased, allowing greater coverage and shorter operation time, which in turn reduces costs. This is demonstrated by an analytical assessment, simulations, and real–world experiments on a group of platforms equipped with low–cost, off–the–shelf sensors.