Showing papers by "Ioannis Pitas published in 2023"

Fast Single-Person 2D Human Pose Estimation Using Multi-Task Convolutional Neural Networks

Abstract: This paper presents a novel neural module for enhancing existing fast and lightweight 2D human pose estimation CNNs, in order to increase their accuracy. A baseline stem CNN is augmented by a collateral module, which is tasked to encode global spatial and semantic information and provide it to the stem network during inference. The latter one outputs the final 2D human pose estimations. Since global information encoding is an inherent subtask of 2D human pose estimation, this particular setup allows the stem network to better focus on the local details of the input image and on precisely localizing each body joint, thus increasing overall 2D human pose estimation accuracy. Furthermore, the collateral module is designed to be lightweight, adding negligible runtime computational cost, so that the unified architecture retains the fast execution property of the stem network. Evaluation of the proposed method on public 2D human pose estimation datasets shows that it increases the accuracy of different baseline stem CNNs, while outperforming all competing fast 2D human pose estimation methods.

Fast CNN-Based Single-Person 2D Human Pose Estimation for Autonomous Systems

Abstract: This paper presents a novel Convolutional Neural Network (CNN) architecture for 2D human pose estimation from RGB images that balances between high 2D human pose/skeleton estimation accuracy and rapid inference. Thus, it is suitable for safety-critical embedded AI scenarios in autonomous systems, where computational resources are typically limited and fast execution is often required, but accuracy cannot be sacrificed. The architecture is composed of a shared feature extraction backbone and two parallel heads attached on top of it: one for 2D human body joint regression and one for global human body structure modelling through Image-to-Image Translation (I2I). A corresponding multitask loss function allows training of the unified network for both tasks, through combining a typical 2D body joint regression with a novel I2I term. Along with enhanced information flow between the parallel neural heads via skip synapses, this strategy is able to extract both ample semantic and rich spatial information, while using a less complex CNN; thus it permits fast execution. The proposed architecture is evaluated on public 2D human pose estimation datasets, achieving the best accuracy-speed ratio compared to the state-of-the-art. Additionally, it is evaluated on a pedestrian intention recognition task for self-driving cars, leading to increased accuracy and speed in comparison to competing approaches.

Estimating Crown Biomass in a Multilayered Fir Forest Using Airborne LiDAR Data

Abstract: The estimation of individual biomass components within tree crowns, such as dead branches (DB), needles (NB), and branch biomass (BB), has received limited attention in the scientific literature despite their significant contribution to forest biomass. This study aimed to assess the potential of multispectral LiDAR data for estimating these biomass components in a multi-layered Abies borissi-regis forest. Destructive (i.e., 13) and non-destructive (i.e., 156) field measurements were collected from Abies borisii-regis trees to develop allometric equations for each crown biomass component and enrich the reference data with the non-destructively sampled trees. A set of machine learning regression algorithms, including random forest (RF), support vector regression (SVR) and Gaussian process (GP), were tested for individual-tree-level DB, NB and BB estimation using LiDAR-derived height and intensity metrics for different spectral channels (i.e., green, NIR and merged) as predictors. The results demonstrated that the RF algorithm achieved the best overall predictive performance for DB (RMSE% = 17.45% and R2 = 0.89), NB (RMSE% = 17.31% and R2 = 0.93) and BB (RMSE% = 24.09% and R2 = 0.85) using the green LiDAR channel. This study showed that the tested algorithms, particularly when utilizing the green channel, accurately estimated the crown biomass components of conifer trees, specifically fir. Overall, LiDAR data can provide accurate estimates of crown biomass in coniferous forests, and further exploration of this method’s applicability in diverse forest structures and biomes is warranted.

Exploiting One-Class Classification Optimization Objectives for Increasing Adversarial Robustness

Abstract: This work examines the problem of increasing the robustness of deep neural network-based image classification systems to adversarial attacks, without changing the neural architecture or employ adversarial examples in the learning process. We attribute their famous lack of robustness to the geometric properties of the deep neural network embedding space, derived from standard optimization options, which allow minor changes in the intermediate activation values to trigger dramatic changes to the decision values in the final layer. To counteract this effect, we explore optimization criteria that supervise the distribution of the intermediate embedding spaces, in a class-specific basis, by introducing and leveraging one-class classification objectives. The proposed learning procedure compares favorably to recently proposed training schemes for adversarial robustness in black-box adversarial attack settings.

Efficient Feature Extraction for Non-Maximum Suppression in Visual Person Detection

Abstract: Non-Maximum Suppression (NMS) is a post-processing step in almost every visual object detector, tasked with rapidly pruning the number of overlapping detected candidate rectangular Regions-of-Interest (RoIs) and replacing them with a single, more spatially accurate detection (in pixel coordinates). The common Greedy NMS algorithm suffers from drawbacks, due to the need for careful manual tuning. In visual person detection, most NMS methods typically suffer when analyzing crowded scenes with high levels of in-between occlusions. This paper proposes a modification on a deep neural architecture for NMS, suitable for such cases and capable of efficiently cooperating with recent neural object detectors. The method approaches the NMS problem as a rescoring task, aiming to ideally assign precisely one detection per object. The proposed modification exploits the extraction of RoI representations, semantically capturing the region’s visual appearance, from information-rich feature maps computed by the detector’s intermediate layers. Experimental evaluation on two common public person detection datasets shows improved accuracy against competing methods, with acceptable inference speed.

Object-Based Burned Area Mapping with Extreme Gradient Boosting Using Sentinel-2 Imagery

Abstract: The Sentinel-2 satellites are providing an unparalleled wealth of high-resolution remotely sensed information with a short revisit cycle, which is ideal for mapping burned areas both accurately and timely. This paper proposes an automated methodology for mapping burn scars using pairs of Sentinel-2 imagery, exploiting the state-of-the-art eXtreme Gradient Boosting (XGB) machine learning framework. A large database of 64 reference wildfire perimeters in Greece from 2016 to 2019 is used to train the classifier. An empirical methodology for appropriately sampling the training patterns from this database is formulated, which guarantees the effectiveness of the approach and its computational efficiency. A difference (pre-fire minus post-fire) spectral index is used for this purpose, upon which we appropriately identify the clear and fuzzy value ranges. To reduce the data volume, a super-pixel segmentation of the images is also employed, implemented via the QuickShift algorithm. The cross-validation results showcase the effectiveness of the proposed algorithm, with the average commission and omission errors being 9% and 2%, respectively, and the average Matthews correlation coefficient (MCC) equal to 0.93.

Towards an Integrated Approach to Wildfire Risk Assessment: When, Where, What and How May the Landscapes Burn

Abstract: This paper presents a review of concepts related to wildfire risk assessment, including the determination of fire ignition and propagation (fire danger), the extent to which fire may spatially overlap with valued assets (exposure), and the potential losses and resilience to those losses (vulnerability). This is followed by a brief discussion of how these concepts can be integrated and connected to mitigation and adaptation efforts. We then review operational fire risk systems in place in various parts of the world. Finally, we propose an integrated fire risk system being developed under the FirEUrisk European project, as an example of how the different risk components (including danger, exposure and vulnerability) can be generated and combined into synthetic risk indices to provide a more comprehensive wildfire risk assessment, but also to consider where and on what variables reduction efforts should be stressed and to envisage policies to be better adapted to future fire regimes. Climate and socio-economic changes entail that wildfires are becoming even more a critical environmental hazard; extreme fires are observed in many areas of the world that regularly experience fire, yet fire activity is also increasing in areas where wildfires were previously rare. To mitigate the negative impacts of fire, those responsible for managing risk must leverage the information available through the risk assessment process, along with an improved understanding on how the various components of risk can be targeted to improve and optimize the many strategies for mitigation and adaptation to an increasing fire risk.

Above-Ground Forest Biomass Estimation using Multispectral LiDAR Data in a Multilayered Coniferous Forest

Abstract: Above-ground biomass and carbon stock are fundamental components of the global carbon cycle, essential for climate change mitigation. Remote sensing data can provide timely and accurate estimates of various forest attributes, especially over large and remote forested areas. The objective of this research was to investigate the potential of multispectral LiDAR data for estimating the stem biomass (SB) and total biomass (TB) in a multi-layered fir forest using an Edge-tree corrected Area Based Approach (EABA). Subsequently, a Random Forest (RF) regression analysis was performed to develop SB and TB predictive models using LiDAR-derived height metrics. Two RF models were produced and evaluated in terms of their predictive performance. Overall, our work demonstrates the capability of multispectral LiDAR data to provide reliable SB and TB estimates in a complex structured forest, contributing significantly to sustainable forest management.