High-resolution RGB Images Research Articles

Learning to Reconstruct and Understand Indoor Scenes From Sparse Views.

This paper proposes a new method for simultaneous 3D reconstruction and semantic segmentation for indoor scenes. Unlike existing methods that require recording a video using a color camera and/or a depth camera, our method only needs a small number of (e.g., 3~5) color images from uncalibrated sparse views, which significantly simplifies data acquisition and broadens applicable scenarios. To achieve promising 3D reconstruction from sparse views with limited overlap, our method first recovers the depth map and semantic information for each view, and then fuses the depth maps into a 3D scene. To this end, we design an iterative deep architecture, named IterNet, to estimate the depth map and semantic segmentation alternately. To obtain accurate alignment between views with limited overlap, we further propose a joint global and local registration method to reconstruct a 3D scene with semantic information. We also make available a new indoor synthetic dataset, containing photorealistic high-resolution RGB images, accurate depth maps and pixel-level semantic labels for thousands of complex layouts. Experimental results on public datasets and our dataset demonstrate that our method achieves more accurate depth estimation, smaller semantic segmentation errors, and better 3D reconstruction results over state-of-the-art methods.

Simultaneous hyperspectral image super-resolution and geometric alignment with a hybrid camera system

Existing snapshot hyperspectral cameras usually suffer from limited spatial resolution in order to maintain reasonable temporal and spectral resolution. In contrast, the spatial resolution of commercial RGB cameras is quite high. Therefore, many methods have been developed for hyperspectral image (HSI) super-resolution by fusing a high resolution RGB image and a low resolution HSI. These methods have been extensively evaluated by using simulated image pair with exact geometric alignment. However, the effect of misalignment on these methods, which always arises in hybrid camera system, has rarely been investigated. In this paper, we present an effective approach for simultaneous HSI super-resolution and geometric alignment of the image pair with drastically contrasting spatial resolution. Besides, we also conduct a systematic evaluation of the misalignment effect on five state-of-the-art hybrid HSI super-resolution methods under five different geometric transformations and three benchmark datasets. Experimental results on both synthetic data and real images show that the proposed method outperforms the current state-of-the-art HSI super-resolution methods with a misaligned hybrid camera system in terms of both objective metric and subjective visual quality.

Spectral Representation vis Data-Guided Sparsity for Hyperspectral Image Super-Resolution†.

Hyperspectral imaging is capable of acquiring the rich spectral information of scenes and has great potential for understanding the characteristics of different materials in many applications ranging from remote sensing to medical imaging. However, due to hardware limitations, the existed hyper-/multi-spectral imaging devices usually cannot obtain high spatial resolution. This study aims to generate a high resolution hyperspectral image according to the available low resolution hyperspectral and high resolution RGB images. We propose a novel hyperspectral image superresolution method via non-negative sparse representation of reflectance spectra with a data guided sparsity constraint. The proposed method firstly learns the hyperspectral dictionary from the low resolution hyperspectral image and then transforms it into the RGB one with the camera response function, which is decided by the physical property of the RGB imaging camera. Given the RGB vector and the RGB dictionary, the sparse representation of each pixel in the high resolution image is calculated with the guidance of a sparsity map, which measures pixel material purity. The sparsity map is generated by analyzing the local content similarity of a focused pixel in the available high resolution RGB image and quantifying the spectral mixing degree motivated by the fact that the pixel spectrum of a pure material should have sparse representation of the spectral dictionary. Since the proposed method adaptively adjusts the sparsity in the spectral representation based on the local content of the available high resolution RGB image, it can produce more robust spectral representation for recovering the target high resolution hyperspectral image. Comprehensive experiments on two public hyperspectral datasets and three real remote sensing images validate that the proposed method achieves promising performances compared to the existing state-of-the-art methods.

Deep Learning Approaches for the Mapping of Tree Species Diversity in a Tropical Wetland Using Airborne LiDAR and High-Spatial-Resolution Remote Sensing Images

The monitoring of tree species diversity is important for forest or wetland ecosystem service maintenance or resource management. Remote sensing is an efficient alternative to traditional field work to map tree species diversity over large areas. Previous studies have used light detection and ranging (LiDAR) and imaging spectroscopy (hyperspectral or multispectral remote sensing) for species richness prediction. The recent development of very high spatial resolution (VHR) RGB images has enabled detailed characterization of canopies and forest structures. In this study, we developed a three-step workflow for mapping tree species diversity, the aim of which was to increase knowledge of tree species diversity assessment using deep learning in a tropical wetland (Haizhu Wetland) in South China based on VHR-RGB images and LiDAR points. Firstly, individual trees were detected based on a canopy height model (CHM, derived from LiDAR points) by the local-maxima-based method in the FUSION software (Version 3.70, Seattle, USA). Then, tree species at the individual tree level were identified via a patch-based image input method, which cropped the RGB images into small patches (the individually detected trees) based on the tree apexes detected. Three different deep learning methods (i.e., AlexNet, VGG16, and ResNet50) were modified to classify the tree species, as they can make good use of the spatial context information. Finally, four diversity indices, namely, the Margalef richness index, the Shannon–Wiener diversity index, the Simpson diversity index, and the Pielou evenness index, were calculated from the fixed subset with a size of 30 × 30 m for assessment. In the classification phase, VGG16 had the best performance, with an overall accuracy of 73.25% for 18 tree species. Based on the classification results, mapping of tree species diversity showed reasonable agreement with field survey data (R2Margalef = 0.4562, root-mean-square error RMSEMargalef = 0.5629; R2Shannon–Wiener = 0.7948, RMSEShannon–Wiener = 0.7202; R2Simpson = 0.7907, RMSESimpson = 0.1038; and R2Pielou = 0.5875, RMSEPielou = 0.3053). While challenges remain for individual tree detection and species classification, the deep-learning-based solution shows potential for mapping tree species diversity.

Hyperspectral Super-Resolution Technique Using Histogram Matching and Endmember Optimization

In most hyperspectral super-resolution (HSR) methods, which are techniques used to improve the resolution of hyperspectral images (HSIs), the HSI and the target RGB image are assumed to have identical fields of view. However, because implementing these identical fields of view is difficult in practical applications, in this paper, we propose a HSR method that is applicable when an HSI and a target RGB image have different spatial information. The proposed HSR method first creates a low-resolution RGB image from a given HSI. Next, a histogram matching is performed on a high-resolution RGB image and a low-resolution RGB image obtained from an HSI. Finally, the proposed method optimizes endmember abundance of the high-resolution HSI towards the histogram-matched high-resolution RGB image. The entire procedure is evaluated using an open HSI dataset, the Harvard dataset, by adding spatial mismatch to the dataset. The spatial mismatch is implemented by shear transformation and cutting off the upper and left sides of the target RGB image. The proposed method achieved a lower error rate across the entire dataset, confirming its capability for super-resolution using images that have different fields of view.

Maize Canopy Temperature Extracted From UAV Thermal and RGB Imagery and Its Application in Water Stress Monitoring.

To identify drought-tolerant crop cultivars or achieve a balance between water use and yield, accurate measurements of crop water stress are needed. In this study, the canopy temperature (Tc) of maize at the late vegetative stage was extracted from high-resolution red–green–blue (RGB, 1.25 cm) and thermal (7.8 cm) images taken by an unmanned aerial vehicle (UAV). To reduce the number of parameters for crop water stress monitoring, four simple methods that require only Tc were identified: Tc, degrees above non-stress, standard deviation of Tc, and variation coefficient of Tc. The ground-truth temperatures obtained using a handheld infrared thermometer were used to calibrate the temperature obtained from the UAV thermal images and to evaluate the Tc extraction results. Measured leaf stomatal conductance values were used to evaluate the performance of the four Tc-based crop water stress indicators. The results showed a strong correlation between ground-truth Tc and Tc extracted by the red–green ratio index (RGRI)-Otsu method proposed in this study, with a coefficient of determination of 0.94 (n = 15) and root mean square error value of 0.7°C. The RGRI-Otsu method was most accurate for estimating temperatures around 32.9°C, but the magnitude of residuals increased above and below this value. This phenomenon may be attributable to changes in canopy cover (leaf curling) under water stress, resulting in changes in the proportion of exposed sunlit soil in UAV thermal orthophotographs. Therefore, to improve the accuracy of maize canopy detection and extraction, optimal methods and better strategies for eliminating mixed pixels are needed. This study demonstrates the potential of using high-resolution UAV RGB images to supplement UAV thermal images for the accurate extraction of maize Tc.

Evaluating the Potential of Semi-Automated Image Analysis for Delimiting Soil and Sediment Layers

ABSTRACTEstablished methods for delineating anthropogenic and natural strata during fieldwork are based on the visual and tactile perception of excavators. Modern image analysis techniques can help to ensure objectivity and reproducibility when documenting sections and plana. Within this study we examine the unsupervised classification of digital images as a technique for delimiting layers and identifying stratigraphic features. Assessing the potential of this approach, we exemplarily captured soil profiles with high-contrast stratigraphy, located in the area of a historical vineyard (Brandenburg, Germany). Reproducible analyses were carried out using open-source software, allowing for the future advancement of the methodology utilized and providing a basis for the analysis of more complex stratigraphic sequences. We compare clustering results of high-resolution RGB and hyperspectral images (470–830 nm, 37 bands). Multiple pre-processing and processing steps are carried out to evaluate their influence. Our results render the semi-automatic analysis of RGB images helpful for stratigraphic interpretation.

Thermal Airborne Optical Sectioning

We apply a multi-spectral (RGB and thermal) camera drone for synthetic aperture imaging to computationally remove occluding vegetation for revealing hidden objects, as required in archeology, search-and-rescue, animal inspection, and border control applications. The radiated heat signal of strongly occluded targets, such as a human bodies hidden in dense shrub, can be made visible by integrating multiple thermal recordings from slightly different perspectives, while being entirely invisible in RGB recordings or unidentifiable in single thermal images. We collect bits of heat radiation through the occluder volume over a wide synthetic aperture range and computationally combine them to a clear image. This requires precise estimation of the drone’s position and orientation for each capturing pose, which is supported by applying computer vision algorithms on the high resolution RGB images.

Quantifying Leaf Phenology of Individual Trees and Species in a Tropical Forest Using Unmanned Aerial Vehicle (UAV) Images

Tropical forests exhibit complex but poorly understood patterns of leaf phenology. Understanding species- and individual-level phenological patterns in tropical forests requires datasets covering large numbers of trees, which can be provided by Unmanned Aerial Vehicles (UAVs). In this paper, we test a workflow combining high-resolution RGB images (7 cm/pixel) acquired from UAVs with a machine learning algorithm to monitor tree and species leaf phenology in a tropical forest in Panama. We acquired images for 34 flight dates over a 12-month period. Crown boundaries were digitized in images and linked with forest inventory data to identify species. We evaluated predictions of leaf cover from different models that included up to 14 image features extracted for each crown on each date. The models were trained and tested with visual estimates of leaf cover from 2422 images from 85 crowns belonging to eight species spanning a range of phenological patterns. The best-performing model included both standard color metrics, as well as texture metrics that quantify within-crown variation, with r2 of 0.84 and mean absolute error (MAE) of 7.8% in 10-fold cross-validation. In contrast, the model based only on the widely-used Green Chromatic Coordinate (GCC) index performed relatively poorly (r2 = 0.52, MAE = 13.6%). These results highlight the utility of texture features for image analysis of tropical forest canopies, where illumination changes may diminish the utility of color indices, such as GCC. The algorithm successfully predicted both individual-tree and species patterns, with mean r2 of 0.82 and 0.89 and mean MAE of 8.1% and 6.0% for individual- and species-level analyses, respectively. Our study is the first to develop and test methods for landscape-scale UAV monitoring of individual trees and species in diverse tropical forests. Our analyses revealed undescribed patterns of high intraspecific variation and complex leaf cover changes for some species.

Weakly supervised pairwise Frank–Wolfe algorithm to recognize a sequence of human actions in RGB-D videos

Human activity recognition is an attractive subject in machine vision which is applicable to intelligent living environments. Daily human activities consist of several actions, in which the boundaries of these actions are different among individuals. The existence of Kinect cameras and RGB-D images, which record joints data, high-resolution RGB and depth images, has improved human action recognition. In this research, actions are recognized by applying the action’s order in the weakly supervised and semi-supervised learning model and extracting the RGB-D data feature. Frank–Wolfe algorithm, which is a constrained convex optimization algorithm, and the pairwise Frank–Wolfe algorithm, which is a developed model of the Frank–Wolfe algorithm, are used as learning models. The evaluation of the proposed method was carried out on the Watch-n-Patch database. The results show a good performance of the proposed method.

RESNET-BASED TREE SPECIES CLASSIFICATION USING UAV IMAGES

Abstract. Tree species classification at individual tree level is a challenging problem in forest management. Deep learning, a cutting-edge technology evolved from Artificial Intelligence, was seen to outperform other techniques when it comes to complex problems such as image classification. In this work, we present a novel method to classify forest tree species through high resolution RGB images acquired with a simple consumer grade camera mounted on a UAV platform using Residual Neural Networks. We used UAV RGB images acquired over three years that varied in numerous acquisition parameters such as season, time, illumination and angle to train the neural network. To begin with, we have experimented with limited data towards the identification of two pine species namely red pine and white pine from the rest of the species. We performed two experiments, first with the images from all three acquisition years and the second with images from only one acquisition year. In the first experiment, we obtained 80% classification accuracy when the trained network was tested on a distinct set of images and in the second experiment, we obtained 51% classification accuracy. As a part of this work, a novel dataset of high-resolution labelled tree species is generated that can be used to conduct further studies involving deep neural networks in forestry.

DETECTING RUMEX OBTUSIFOLIUS WEED PLANTS IN GRASSLANDS FROM UAV RGB IMAGERY USING DEEP LEARNING

Abstract. Broad-leaved dock (Rumex obtusifolius) is a fast growing and spreading weed and is one of the most common weeds in production grasslands in the Netherlands. The heavy occurrence, fast growth and negative environmental-agricultural impact makes Rumex a species important to control. Current control is done directly in the field by mechanical or chemical actuation methods as soon as the plants are found in situ by the farmer. In nature conservation areas control is much more difficult because spraying is not allowed. This reduces the amount of grass and its quality. Rumex could be rapidly detected using high-resolution RGB images obtained from a UAV and optimize the plant control practices in wide nature conservation areas. In this paper, a novel approach for Rumex detection from orthomosaics obtained using a commercial available quadrotor (DJI Phantom 3 PRO) is proposed. The results obtained shown that Rumex can be detected up to 90% from a 6 mm/pixel ortho-mosaic generated from an aerial survey and using deep learning.

A Low-Cost and Unsupervised Image Recognition Methodology for Yield Estimation in a Vineyard.

Yield prediction is a key factor to optimize vineyard management and achieve the desired grape quality. Classical yield estimation methods, which consist of manual sampling within the field on a limited number of plants before harvest, are time-consuming and frequently insufficient to obtain representative yield data. Non-invasive machine vision methods are therefore being investigated to assess and implement a rapid grape yield estimate tool. This study aimed at an automated estimation of yield in terms of cluster number and size from high resolution RGB images (20 MP) taken with a low-cost UAV platform in representative zones of the vigor variability within an experimental vineyard. The flight campaigns were conducted in different light conditions and canopy cover levels for 2017 and 2018 crop seasons. An unsupervised recognition algorithm was applied to derive cluster number and size, which was used for estimating yield per vine. The results related to the number of clusters detected in different conditions, and the weight estimation for each vigor zone are presented. The segmentation results in cluster detection showed a performance of over 85% in partially leaf removal and full ripe condition, and allowed grapevine yield to be estimated with more than 84% of accuracy several weeks before harvest. The application of innovative technologies in field-phenotyping such as UAV, high-resolution cameras and visual computing algorithms enabled a new methodology to assess yield, which can save time and provide an accurate estimate compared to the manual method.

Phenotyping of Plant Biomass and Performance Traits Using Remote Sensing Techniques in Pea (Pisum sativum, L.).

Field pea cultivars are constantly improved through breeding programs to enhance biotic and abiotic stress tolerance and increase seed yield potential. In pea breeding, the Above Ground Biomass (AGBM) is assessed due to its influence on seed yield, canopy closure, and weed suppression. It is also the primary yield component for peas used as a cover crop and/or grazing. Measuring AGBM is destructive and labor-intensive process. Sensor-based phenotyping of such traits can greatly enhance crop breeding efficiency. In this research, high resolution RGB and multispectral images acquired with unmanned aerial systems were used to assess phenotypes in spring and winter pea breeding plots. The Green Red Vegetation Index (GRVI), Normalized Difference Vegetation Index (NDVI), Normalized Difference Red Edge Index (NDRE), plot volume, canopy height, and canopy coverage were extracted from RGB and multispectral information at five imaging times (between 365 to 1948 accumulated degree days/ADD after 1 May) in four winter field pea experiments and at three imaging times (between 1231 to 1648 ADD) in one spring field pea experiment. The image features were compared to ground-truth data including AGBM, lodging, leaf type, days to 50% flowering, days to physiological maturity, number of the first reproductive node, and seed yield. In two of the winter pea experiments, a strong correlation between image features and seed yield was observed at 1268 ADD (flowering). An increase in correlation between image features with the phenological traits such as days to 50% flowering and days to physiological maturity was observed at about 1725 ADD in these winter pea experiments. In the spring pea experiment, the plot volume estimated from images was highly correlated with ground truth canopy height (r = 0.83) at 1231 ADD. In two other winter pea experiments and the spring pea experiment, the GRVI and NDVI features were significantly correlated with AGBM at flowering. When selected image features were used to develop a least absolute shrinkage and selection operator model for AGBM estimation, the correlation coefficient between the actual and predicted AGBM was 0.60 and 0.84 in the winter and spring pea experiments, respectively. A SPOT-6 satellite image (1.5 m resolution) was also evaluated for its applicability to assess biomass and seed yield. The image features extracted from satellite imagery showed significant correlation with seed yield in two winter field pea experiments, however, the trend was not consistent. In summary, the study supports the potential of using unmanned aerial system-based imaging techniques to estimate biomass and crop performance in pea breeding programs.

Extended Super Resolution of Hyperspectral Images via Non-negative Sparse Coding

Often high resolution (HR) RGB images generated by sparse sampling of the visible spectrum fail to produce differentiable modality for computer vision tasks, Hence computer vision tasks have to rely on gross structures in an image like corners, edges etc. instead of just the recorded reflectance of objects or materials at each pixel in a scene. In contrast to RGB, hyperspectral imaging allows pixels to record reflectance of the scene over multiple contiguous bands, which results in rich differentiable modalities. However, hyperspectral imaging, despite having a growing number of applications from agriculture, surveillance, mineralogy, food processing to eye care, is hitherto restricted to low spatial resolution imaging due to sensor hardware limitations. In this paper, we propose a hyperspectral super resolution technique to produce a high resolution (HR) hyperspectral image with a spectral support of 400–1020 nm from a low resolution (LR) hyperspectral image of the same spectral support and a high resolution multispectral (RGB) image with reduced spectral support of 400–700 nm. In the first step, we generate a HR hyperspectral prior by estimating HR hyperspectral band images in the spectral support of 400–700 nm by detail transfer and alternating iterative minimization. In the next step, we use the generated HR prior to further estimate the HR hyperspectral images for 710–1020 nm bands by learning a non-negative dictionary of reflectance spectra signatures of all the materials present in the scene from the LR hyperspectral image with spectral support in 400–1020 nm. With the estimated HR hyperspectral prior and learned dictionary, we predict the non-negative sparse codes for HR hyperspectral band images in the band of 710–1020 nm.

Detection of fallen logs from high-resolution UAV images

Background: High-resolution images from unmanned aerial vehicles (UAVs) can be used to describe the state of forests at regular time periods in a cost-effective manner. The purpose of this study was to assess the performance of a line template matching algorithm, the Hough transformation, for detecting fallen logs from UAV-based high-resolution RGB images. The suggested methodology does not aim to replace any known aerial method for log detection, rather it is more oriented to the detection of fallen logs in open forest stands with a high percentage of log visibility and straightness.
 Methods: This study describes a line template matching algorithm that can be used for the detection of fallen logs in an automated process. The detection technique was based on object-based image analysis, using both pixel-based and shape descriptors. To determine the actual number of fallen logs, and to compare with the ones predicted by the algorithm, manual visual assessment was used based on six high-resolution orthorectified images. To evaluate if a line matched, we used a voting scheme. The total number of detected fallen logs compared with the actual number of fallen logs based on several accuracy metrics. To evaluate predictive models we tested the cross-validation mean error. Finally, to test how close our results were to chance, we used the Cohen`s Kappa coefficient.
 Results: The detection algorithm found 136 linear objects, of which 92 of them were detected as fallen logs. From the 92 detected fallen logs, 86 were correctly predicted by the algorithm and 24 were falsely detected as fallen logs. The calculated amount of observed agreement was equal to 0.78, whereas the expected agreement by chance was 0.61. Finally, the kappa statistic was 0.44.
 Conclusions: Our methodology had high reliability for detecting fallen logs based on total user‘s accuracy (94.9%), whereas a Kappa of 0.44 indicated there was good agreement between the observed and predicted values. Also, the cross-validation analysis denoted the efficiency of the proposed method with an average error of 16%.

High-Fidelity Depth Upsampling Using the Self-Learning Framework.

This paper presents a depth upsampling method that produces a high-fidelity dense depth map using a high-resolution RGB image and LiDAR sensor data. Our proposed method explicitly handles depth outliers and computes a depth upsampling with confidence information. Our key idea is the self-learning framework, which automatically learns to estimate the reliability of the upsampled depth map without human-labeled annotation. Thereby, our proposed method can produce a clear and high-fidelity dense depth map that preserves the shape of object structures well, which can be favored by subsequent algorithms for follow-up tasks. We qualitatively and quantitatively evaluate our proposed method by comparing other competing methods on the well-known Middlebury 2014 and KITTIbenchmark datasets. We demonstrate that our method generates accurate depth maps with smaller errors favorable against other methods while preserving a larger number of valid points, as we also show that our approach can be seamlessly applied to improve the quality of depth maps from other depth generation algorithms such as stereo matching and further discuss potential applications and limitations. Compared to previous work, our proposed method has similar depth errors on average, while retaining at least 3% more valid depth points.

Multispectral Image Super-Resolution via RGB Image Fusion and Radiometric Calibration.

Multispectral imaging is of wide application for its capability in acquiring the spectral information of scenes. Due to hardware limitation, multispectral imaging device usually cannot achieve high-spatial resolution. To address the issue, this paper proposes a multispectral image super-resolution algorithm, referred as SRIF, by fusing the low-resolution multispectral image and the high-resolution (HR) RGB image. It deals with the general circumstance that image intensity is linear to scene radiance for multispectral imaging devices while is nonlinear and unknown for most RGB cameras. The SRIF algorithm first solves the inverse camera response function and a spectral sensitivity function of RGB camera, and establishes the linear relationship between multispectral and RGB images. Then the unknown HR multispectral image is efficiently reconstructed according to the linear image degradation models. Meanwhile, the edge structure of the reconstructed HR multispectral image is kept in accordance with that of the RGB image using a weighted total variation regularizer. The effectiveness of the SRIF algorithm is evaluated on both public datasets and our image set. Experimental results validate that the SRIF algorithm outperforms the state-of-the-arts in terms of both reconstruction accuracy and computational efficiency.

Image based classification of slums, built-up and non-built-up areas in Kalyan and Bangalore, India

ABSTRACT Slums, characterized by sub-standard housing conditions, are a common in fast growing Asian cities. However, reliable and up-to-date information on their locations and development dynamics is scarce. Despite numerous studies, the task of delineating slum areas remains a challenge and no general agreement exists about the most suitable method for detecting or assessing detection performance. In this paper, standard computer vision methods – Bag of Visual Words framework and Speeded-Up Robust Features have been applied for image-based classification of slum and non-slum areas in Kalyan and Bangalore, India, using very high resolution RGB images. To delineate slum areas, image segmentation is performed as pixel-level classification for three classes: Slums, Built-up and Non-Built-up. For each of the three classes, image tiles were randomly selected using ground truth observations. A multi-class support vector machine classifier has been trained on 80% of the tiles and the remaining 20% were used for testing. The final image segmentation has been obtained by classification of every 10th pixel followed by a majority filtering assigning classes to all remaining pixels. The results demonstrate the ability of the method to map slums with very different visual characteristics in two very different Indian cities.

Ear density estimation from high resolution RGB imagery using deep learning technique

Wheat ear density estimation is an appealing trait for plant breeders. Current manual counting is tedious and inefficient. In this study we investigated the potential of convolutional neural networks (CNNs) to provide accurate ear density using nadir high spatial resolution RGB images. Two different approaches were investigated, either using the Faster-RCNN state-of-the-art object detector or with the TasselNet local count regression network. Both approaches performed very well (rRMSE≈6%) when applied over the same conditions as those prevailing for the calibration of the models. However, Faster-RCNN was more robust when applied to a dataset acquired at a later stage with ears and background showing a different aspect because of the higher maturity of the plants. Optimal spatial resolution for Faster-RCNN was around 0.3 mm allowing to acquire RGB images from a UAV platform for high-throughput phenotyping of large experiments. Comparison of the estimated ear density with in-situ manual counting shows reasonable agreement considering the relatively small sampling area used for both methods. Faster-RCNN and in-situ counting had high and similar heritability (H²≈85%), demonstrating that ear density derived from high resolution RGB imagery could replace the traditional counting method.