NVIDIA Jetson TX2 Research Articles

Removing nonrigid refractive distortions for underwater images using an attention-based deep neural network

Refractive distortions in underwater images usually occur when these images are captured through a dynamic refractive water surface, such as unmanned aerial vehicles capturing shallow underwater scenes from the surface of water or autonomous underwater vehicles observing floating platforms in the air. We propose an end-to-end deep neural network for learning to restore real scene images for removing refractive distortions. This network adopts an encoder-decoder architecture with a specially designed attention module. The use of the attention image and the distortion field generated by the proposed deep neural network can restore the exact distorted areas in more detail. Qualitative and quantitative experimental results show that the proposed framework effectively eliminates refractive distortions and refines image details. We also test the proposed framework in practical applications by embedding it into an NVIDIA JETSON TX2 platform, and the results demonstrate the practical value of the proposed framework.

A Lightweight Insulator Defect Detection Model Based on Drone Images

With the continuous development and construction of new power systems, using drones to inspect the condition of transmission line insulators has become an inevitable trend. To facilitate the deployment of drone hardware equipment, this paper proposes IDD-YOLO (Insulator Defect Detection-YOLO), a lightweight insulator defect detection model. Initially, the backbone network of IDD-YOLO employs GhostNet for feature extraction. However, due to the limited feature extraction capability of GhostNet, we designed a lightweight attention mechanism called LCSA (Lightweight Channel-Spatial Attention), which is combined with GhostNet to capture features more comprehensively. Secondly, the neck network of IDD-YOLO utilizes PANet for feature transformation and introduces GSConv and C3Ghost convolution modules to reduce redundant parameters and lighten the network. The head network employs the YOLO detection head, incorporating the EIOU loss function and Mish activation function to optimize the speed and accuracy of insulator defect detection. Finally, the model is optimized using TensorRT and deployed on the NVIDIA Jetson TX2 NX mobile platform to test the actual inference speed of the model. The experimental results demonstrate that the model exhibits outstanding performance on both the proprietary ID-2024 insulator defect dataset and the public SFID insulator dataset. After optimization with TensorRT, the actual inference speed of the IDD-YOLO model reached 20.83 frames per second (FPS), meeting the demands for accurate and real-time inspection of insulator defects by drones.

An Intelligent Bait Delivery Control Method for Flight Vehicle Evasion Based on Reinforcement Learning

During aerial combat, when an aircraft is facing an infrared air-to-air missile strike, infrared baiting technology is an important means of penetration, and the strategy of effective delivery of infrared bait is critical. To address this issue, this study proposes an improved deep deterministic policy gradient (DDPG) algorithm-based intelligent bait-dropping control method. Firstly, by modeling the relative motion between aircraft, bait, and incoming missiles, the Markov decision process of aircraft-bait-missile infrared effect was constructed with visual distance and line of sight angle as states. Then, the DDPG algorithm was improved by means of pre-training and classification sampling. Significantly, the infrared bait-dropping decision network was trained through interaction with the environment and iterative learning, which led to the development of the bait-dropping strategy. Finally, the corresponding environment was transferred to the Nvidia Jetson TX2 embedded platform for comparative testing. The simulation results showed that the convergence speed of this method was 46.3% faster than the traditional DDPG algorithm. More importantly, it was able to generate an effective bait-throwing strategy, enabling the aircraft to successfully evade the attack of the incoming missile. The strategy instruction generation time is only about 2.5 ms, giving it the ability to make online decisions.

A lightweight and accurate recognition algorithm of pointer meter based on improved Deeplabv3+ for inspection robots

Pointer meter is widely utilized in the fields of modern industry. Nowadays, intelligent inspection robots are gradually employed in place of labors for inspection task. Pointer meter recognition is one of the most important tasks of inspection robots. This article presents a lightweight pointer meter recognition algorithm, which is suitable for deployment on inspection robots. First, pointer meter is identified by pruned YOLOv5. Afterwards, the dial and the pointer are segmented through improved Deeplabv3+ in which the JPU and depthwise separable convolutions are utilized in lieu of dilated convolutions. After perspective transform and central-line extraction of pointer, the reading of the pointer meter can be determined using the angle method. Experimental results verify the FPS of pruned YOLOv5 improves 13.18% compared to original YOLOv5 and the FPS of improved Deeplabv3+ improves 45.74% compared to original Deeplabv3+. Additionally, the reading accuracy of the algorithm is 98.40% and average fiducial error is 0.32%, which indicate good accuracy. The relative standard deviation of reading is less than 1.4%, which indicates good stability of proposed algorithm. This study proposes a lightweight and accurate pointer meter recognition algorithm based on improved Deeplabv3+, the algorithm is ported to NVIDIA Jetson TX2 NX to verify its stability and accuracy.

AI-Driven Computer Vision Detection of Cotton in Corn Fields Using UAS Remote Sensing Data and Spot-Spray Application

To effectively combat the re-infestation of boll weevils (Anthonomus grandis L.) in cotton fields, it is necessary to address the detection of volunteer cotton (VC) plants (Gossypium hirsutum L.) in rotation crops such as corn (Zea mays L.) and sorghum (Sorghum bicolor L.). The current practice involves manual field scouting at the field edges, which often leads to the oversight of VC plants growing in the middle of fields alongside corn and sorghum. As these VC plants reach the pinhead squaring stage (5–6 leaves), they can become hosts for boll weevil pests. Consequently, it becomes crucial to detect, locate, and accurately spot-spray these plants with appropriate chemicals. This paper focuses on the application of YOLOv5m to detect and locate VC plants during the tasseling (VT) growth stage of cornfields. Our results demonstrate that VC plants can be detected with a mean average precision (mAP) of 79% at an Intersection over Union (IoU) of 50% and a classification accuracy of 78% on images sized 1207 × 923 pixels. The average detection inference speed is 47 frames per second (FPS) on the NVIDIA Tesla P100 GPU-16 GB and 0.4 FPS on the NVIDIA Jetson TX2 GPU, which underscores the relevance and impact of detection speed on the feasibility of real-time applications. Additionally, we show the application of a customized unmanned aircraft system (UAS) for spot-spray applications through simulation based on the developed computer vision (CV) algorithm. This UAS-based approach enables the near-real-time detection and mitigation of VC plants in corn fields, with near-real-time defined as approximately 0.02 s per frame on the NVIDIA Tesla P100 GPU and 2.5 s per frame on the NVIDIA Jetson TX2 GPU, thereby offering an efficient management solution for controlling boll weevil pests.

Design and Implementation of Intelligent Robot Control System Integrating Computer Vision and Mechanical Engineering

With the rapid development of science and technology and industry, robot technology has become a key pillar in many fields, especially in industrial manufacturing, robots have been widely used in automated operations. However, in the face of the complex and changeable demands of modern industry, simple mechanized operation can no longer be satisfied, and intelligence has become the inevitable trend of robot technology development. By combining computer vision technology, this study provides robots with powerful environmental awareness and target recognition capabilities, thus achieving higher-level autonomous operation and decision-making. Aiming at the intelligent sorting system, this research adopts XYZ mechanical arm as the execution platform, equipped with TCS34725 color sensor to identify the color of articles, and integrates Hikvision industrial camera and NVIDIA Jetson TX2 image processor to capture and process image data, so as to achieve accurate positioning of articles. In the aspect of software design, this research realizes the function of accurately sorting different color items from the conveyor belt by combining color recognition algorithm and image processing technology with mechanical engineering control. The test results of the system show that the system has high accuracy and stability in color recognition and object positioning, and the trajectory accuracy of the manipulator meets the requirements, which verifies the realization of the research goal. The intelligent robot control system designed in this paper not only improves the level of industrial automation and reduces labor costs, but also broadens the application fields of robots in complex environments.

Investigating visual navigation using spiking neural network models of the insect mushroom bodies.

Ants are capable of learning long visually guided foraging routes with limited neural resources. The visual scene memory needed for this behaviour is mediated by the mushroom bodies; an insect brain region important for learning and memory. In a visual navigation context, the mushroom bodies are theorised to act as familiarity detectors, guiding ants to views that are similar to those previously learned when first travelling along a foraging route. Evidence from behavioural experiments, computational studies and brain lesions all support this idea. Here we further investigate the role of mushroom bodies in visual navigation with a spiking neural network model learning complex natural scenes. By implementing these networks in GeNN-a library for building GPU accelerated spiking neural networks-we were able to test these models offline on an image database representing navigation through a complex outdoor natural environment, and also online embodied on a robot. The mushroom body model successfully learnt a large series of visual scenes (400 scenes corresponding to a 27 m route) and used these memories to choose accurate heading directions during route recapitulation in both complex environments. Through analysing our model's Kenyon cell (KC) activity, we were able to demonstrate that KC activity is directly related to the respective novelty of input images. Through conducting a parameter search we found that there is a non-linear dependence between optimal KC to visual projection neuron (VPN) connection sparsity and the length of time the model is presented with an image stimulus. The parameter search also showed training the model on lower proportions of a route generally produced better accuracy when testing on the entire route. We embodied the mushroom body model and comparator visual navigation algorithms on a Quanser Q-car robot with all processing running on an Nvidia Jetson TX2. On a 6.5 m route, the mushroom body model had a mean distance to training route (error) of 0.144 ± 0.088 m over 5 trials, which was performance comparable to standard visual-only navigation algorithms. Thus, we have demonstrated that a biologically plausible model of the ant mushroom body can navigate complex environments both in simulation and the real world. Understanding the neural basis of this behaviour will provide insight into how neural circuits are tuned to rapidly learn behaviourally relevant information from complex environments and provide inspiration for creating bio-mimetic computer/robotic systems that can learn rapidly with low energy requirements.

A photogrammetric approach for real‐time visual SLAM applied to an omnidirectional system

AbstractThe problem of sequential estimation of the exterior orientation of imaging sensors and the three‐dimensional environment reconstruction in real time is commonly known as visual simultaneous localisation and mapping (vSLAM). Omnidirectional optical sensors have been increasingly used in vSLAM solutions, mainly for providing a wider view of the scene, allowing the extraction of more features. However, dealing with unmodelled points in the hyperhemispherical field poses challenges, mainly due to the complex lens geometry entailed in the image formation process. To address these challenges, the use of rigorous photogrammetric models that appropriately handle the geometry of fisheye lens cameras can overcome these challenges. Thus, this study presents a real‐time vSLAM approach for omnidirectional systems adapting ORB‐SLAM with a rigorous projection model (equisolid‐angle). The implementation was conducted on the Nvidia Jetson TX2 board, and the approach was evaluated using hyperhemispherical images captured by a dual‐fisheye camera (Ricoh Theta S) embedded into a mobile backpack platform. The trajectory covered a distance of 140 m, with the approach demonstrating accuracy better than 0.12 m at the beginning and achieving metre‐level accuracy at the end of the trajectory. Additionally, we compared the performance of our proposed approach with a generic model for fisheye lens cameras.

Convolution technique for focusing of ISAR images

ABSTRACT Defocusing is an undesirable phenomenon in remote sensing. Defocusing arises due to the platform motion irregularity or the target motion. The focusing techniques in the literature either need the imaging parameters to compensate for the motion aberrations or a large dataset of defocused, focused images to model the phase errors. This paper proposes a focusing technique called convolution technique that uses only the blurred image data to focus it. The experimental results on the inverse synthetic aperture (ISAR) dataset show that the proposed technique achieved better entropy and contrast than state-of-the-art methods. The execution time of the proposed technique on NVIDIA Jetson Tx2 device is 182.9 msec. Hence, the convolution technique can be used for real-time applications.

Sugar Beet Seed Classification for Production Quality Improvement by Using YOLO and NVIDIA Artificial Intelligence Boards

AbstractAll inputs are required for excellent and proper crop production, especially seed quality. In this way fewer disease and insect issues, increased seedling germination, uniform plant population and maturity, and better responsiveness to fertilizers and nutrients, leading to higher returns per unit area and profitability, and low labor costs could be possible. Because of this reason, NVIDIA Jetson Nano and TX2 artificial intelligence boards were used to test the efficiency of the YOLOv4 and YOLOv4-tiny models for sugar beet monogerm and multigerm seed classification for better production. YOLOv4-tiny outscored the other model based on FPS with 8.25–8.37 at NVIDIA Jetson Nano, 12.11–12.36 at NVIDIA TX2 artificial intelligence boards with accuracy 81–99% for monogerm seeds, and 89–99% for multigerm seeds at NVIDIA Jetson Nano, 88–99% for monogerm seeds, and 90–99% for multigerm at NVIDIA TX2 accuracy, respectively, implying that the YOLOv4 is more accurate but slow with based on FPS with 1.10–1.21 at NVIDIA Jetson Nano, 2.41–2.43 at NVIDIA TX2 artificial intelligence boards with 95–99% for monogerm seeds and 95–100% for multigerm seeds at NVIDIA Jetson Nano, 92–99% for monogerm seeds and 98–100% for multigerm seeds at NVIDIA TX2, respectively. As a result of the evaluations, NVIDIA Artificial Intelligence cards and YOLO deep learning model will be used effectively in classifying monogerm and multigerm sugar beet seeds, thus reducing seed loss with the help of NVIDIA Artificial Intelligence cards classification.

Efficient object tracking on edge devices with MobileTrack

Object tracking has made significant progress in recent years. However, the state-of-the-art trackers are becoming increasingly heavy and expensive, making their deployment challenging in resource-constrained applications. In this study, we introduce MobileTrack, a visual object tracker that strikes a perfect balance between tracking accuracy and inference speed. Utilizing a novel coordinated perception-aware fusion module and a lightweight prediction head, our proposed methodology outperforms most Siamese trackers on various academic benchmarks in terms of both accuracy and efficiency. When deployed on resource-constrained embedded devices such as NVIDIA Jetson TX2, MobileTrack ensures real-time performance at a speed exceeding 33FPS, while LightTrack only operates at 18FPS. Therefore, MobileTrack holds significant potential to unlock a wide range of practical applications across various industries. MobileTrack is released at here.

Enhancing Facial Recognition in Surveillance Systems through Embedded Super-resolution

This document details the implementation of a sub-pixel convolutional neural network designed to enhance the resolution of face images. The model uses a series of filters to progressively increase the number of pixels, estimating the necessary information for new pixels from the original image and training derived from 22000 synthetic images produced by adversarial neural networks. Within the context of surveillance and related applications, the trained convolutional network exhibits beneficial characteristics. For instance, it can be deployed within a device to achieve higher-resolution images than those the physical camera can produce. This research underscores the feasibility of such a device through the implementation and evaluation of the network on the NVIDIA Jetson TX2 embedded system. The findings demonstrate the model's practicality for real-time surveillance applications and its ability to produce superior-quality images compared to several interpolation methods, as determined by an exhaustive testing process measuring various attributes of the generated images.

A real-time and accurate convolutional neural network for fabric defect detection

As a practical and challenging task, deep learning-based methods have achieved effective results for fabric defect detection, however, most of them mainly target detection accuracy at the expense of detection speed. Therefore, we propose a fabric defect detection method called PEI-YOLOv5. First, Particle Depthwise Convolution (PDConv) is proposed to extract spatial features more efficiently while reducing redundant computations and memory access, reducing model computation and improving detection speed. Second, Enhance-BiFPN(EB) is proposed based on the structure of BiFPN to enhance the attention of spatial and channel feature maps and the fusion of information at different scales. Third, we improve the loss function and propose IN loss, which improves the problem that the original IOU loss is weak in detecting small targets while speeding up the convergence of the model. Finally, five more common types of defects were selected for training in the GuangDong TianChi fabric defect dataset, and using our proposed PEI-YOLOv5 with only 0.2 Giga Floating Point Operations (GFLOPs) increase, the mAP improved by 3.61%, reaching 87.89%. To demonstrate the versatility of PEI-YOLOv5, we additionally evaluated this in the NEU surface defect database, with the mAP of 79.37%. The performance of PEI-YOLOv 5 in these two datasets surpasses the most advanced fabric defect detection methods at present. We deployed the model to the NVIDIA Jetson TX2 embedded development board, and the detection speed reached 31 frames per second (Fps), which can fully meet the speed requirements of real-time detection.

A Binocular Stereo-Imaging-Perception System with a Wide Field-of-View and Infrared- and Visible Light-Dual-Band Fusion.

With the continuous evolution of autonomous driving and unmanned driving systems, traditional limitations such as a limited field-of-view, poor ranging accuracy, and real-time display are becoming inadequate to satisfy the requirements of binocular stereo-perception systems. Firstly, we designed a binocular stereo-imaging-perception system with a wide-field-of-view and infrared- and visible light-dual-band fusion. Secondly we proposed a binocular stereo-perception optical imaging system with a wide field-of-view of 120.3°, which solves the small field-of-view of current binocular stereo-perception systems. Thirdly, For image aberration caused by the wide-field-of-view system design, we propose an ellipsoidal-image-aberration algorithm with a low consumption of memory resources and no loss of field-of-view. This algorithm simultaneously solves visible light and infrared images with an aberration rate of 45% and 47%, respectively. Fourthly, a multi-scale infrared- and visible light-image-fusion algorithm is used, which improves the situational-awareness capabilities of a binocular stereo-sensing system in a scene and enhances image details to improve ranging accuracy. Furthermore, this paper is based on the Taylor model-calibration binocular stereo-sensing system of internal and external parameters for limit correction; the implemented algorithms are integrated into an NVIDIA Jetson TX2 + FPGA hardware framework, enabling near-distance ranging experiments. The fusion-ranging accuracy within 20 m achieved an error of 0.02 m, outperforming both visible light- and infrared-ranging methods. It generates the fusion-ranging-image output with a minimal delay of only 22.31 ms at a frame rate of 50 Hz.

Comprehensive analysis of the heterogeneous computing performance of DNNs under typical frameworks on cloud and edge computing platforms

In this paper, nearly 40 commonly used deep neural network(DNN) models are selected, and their cross-platform and cross-inference frameworks are deeply analysed. The main metrics of accuracy, the total number of model parameters, the computational complexity, the accuracy density, the inference time, the memory consumption and other related parameters are used to measure their performance. The heterogeneous computing experiment is implemented on both the Google Colab cloud computing platform and the Jetson Nano embedded edge computing platform. The obtained performance is compared with that of two previous computing platforms: a workstation equipped with an NVIDIA Titan X Pascal and an embedded system based on an NVIDIA Jetson TX1 board. In addition, on the Jetson Nano embedded edge computing platform, different inference frameworks are investigated to evaluate the inference efficiency of the DNN models. Regression models are established to characterize the variation in the computing performance of different DNN classification algorithms so that the inference results of unknown models can be estimated. ANOVA methods are proposed to quantify the differences between models. The experimental results have important guiding significance for the better selection, deployment and application of DNN models in practice. Codes are available at this https URL https://github.com/Foreverzfy/Model-Test.

Robust Person Identification and Following in a Mobile Robot Based on Deep Learning and Optical Tracking

There is an exciting synergy between deep learning and robotics, combining the perception skills a deep learning system can achieve with the wide variety of physical responses a robot can perform. This article describes an embedded system integrated into a mobile robot capable of identifying and following a specific person reliably based on a convolutional neural network pipeline. In addition, the design incorporates an optical tracking system for supporting the inferences of the neural networks, allowing the determination of the position of a person using an RGB depth camera. The system runs on an NVIDIA Jetson TX2 board, an embedded System-on-Module capable of performing computationally demanding tasks onboard and handling the complexity needed to run a solid tracking and following algorithm. A robotic mobile base with the Jetson TX2 board attached receives velocity orders to move the system toward the target. The proposed approach has been validated on a mobile robotic platform that successfully follows a determined person, relying on the robustness of the combination of deep learning with optical tracking for working efficiently in a real environment.

Research on Intelligent Control Method of Launch Vehicle Landing Based on Deep Reinforcement Learning

A launch vehicle needs to adapt to a complex flight environment during flight, and traditional guidance and control algorithms can hardly deal with multi-factor uncertainties due to the high dependency on control models. To solve this problem, this paper designs a new intelligent flight control method for a rocket based on the deep reinforcement learning algorithm driven by knowledge and data. In this process, the Markov decision process of the rocket landing section is established by designing a reinforcement function with consideration of the combination effect on the return of the terminal constraint of the launch vehicle and the cumulative return of the flight process of the rocket. Meanwhile, to improve the training speed of the landing process of the launch vehicle and to enhance the generalization ability of the model, the strategic neural network model is obtained and trained via the form of a long short-term memory (LSTM) network combined with a full connection layer as a landing guidance strategy network. The proximal policy optimization (PPO) is the training algorithm of reinforcement learning network parameters combined with behavioral cloning (BC) as the reinforcement learning pre-training imitation learning algorithm. Notably, the rocket-borne environment is transplanted to the Nvidia Jetson TX2 embedded platform for the comparative testing and verification of this intelligent model, which is then used to generate real-time control commands for guiding the actual flying and landing process of the rocket. Further, comparisons of the results obtained from convex landing optimization and the proposed method in this work are performed to prove the effectiveness of this proposed method. The simulation results show that the intelligent control method in this work can meet the landing accuracy requirements of the launch vehicle with a fast convergence speed of 84 steps, and the decision time is only 2.5 ms. Additionally, it has the ability of online autonomous decision making as deployed on the embedded platform.

An improved YOLOv5s model for effectively predict sugarcane seed replenishment positions verified by a field re-seeding robot

Sugarcane field re-seeding robot is a promising yield-enhancing technology proposed to solve the seedling absences in sugarcane fields. In this study, in combination with developing the sugarcane field re-seeding robot, an improved YOLOv5s model was proposed to detect sugarcane seedlings and predict seed replenishment positions. Firstly, field images of one-month-old sugarcane seedlings were taken at different light conditions as a dataset. Secondly, the Slim-Neck was introduced to replace the Neck network, which can reduce the complexity of the model while maintaining sufficient accuracy. Thirdly, the Efficient Channel Attention (ECA) module was added to the Backbone network to enhance the model's attention on critical feature information of sugarcane seedlings. Fourthly, the SCYLLA-IoU (SIoU) loss function was introduced to speed up the convergence of the proposed model. Lastly, a method for predicting seed replenishment positions was proposed and verified by the field tests. The experimental results showed that the mean average precision (mAP), precision, and recall of the improved YOLOv5s model were 93.1 %, 92.1 %, and 89.9 %, respectively, and the detection speed was 82 frames per second (FPS), which increased the mAP by 1.5 % and the detection speed by 12.3 % compared to the original YOLOv5s model. In addition, compared with Faster R-CNN, SSD, and YOLOv4-tiny models, the improved YOLOv5s model had a higher accuracy, faster detection speed, and less memory consumption. The field test showed that the real-time detection speed of the improved YOLOv5s model was 23 FPS in Nvidia Jetson TX2. The real-time detection precision of sugarcane seedlings was 97.2 %, and the recall was 86.7 %. The mean relative error between the numbers of seed replenishment positions predicted by the robot and that predicted by the human was 18.7 %. Consequently, the improved YOLOv5s model can efficiently and accurately detect sugarcane seedlings and predict seed replenishment positions. This technology provides valuable visual detection support for the sugarcane field re-seeding robot.

Towards Real-Time On-Drone Pedestrian Tracking in 4K Inputs

Over the past several years, significant progress has been made in object tracking, but challenges persist in tracking objects in high-resolution images captured from drones. Such images usually contain very tiny objects, and the movement of the drone causes rapid changes in the scene. In addition, the computing power of mission computers on drones is often insufficient to achieve real-time processing of deep learning-based object tracking. This paper presents a real-time on-drone pedestrian tracker that takes as the input 4K aerial images. The proposed tracker effectively hides the long latency required for deep learning-based detection (e.g., YOLO) by exploiting both the CPU and GPU equipped in the mission computer. We also propose techniques to minimize detection loss in drone-captured images, including a tracker-assisted confidence boosting and an ensemble for identity association. In our experiments, using real-world inputs captured by drones at a height of 50 m, the proposed method with an NVIDIA Jetson TX2 proves its efficacy by achieving real-time detection and tracking in 4K video streams.

METER: A Mobile Vision Transformer Architecture for Monocular Depth Estimation

Depth estimation is a fundamental knowledge for autonomous systems that need to assess their own state and perceive the surrounding environment. Deep learning algorithms for depth estimation have gained significant interest in recent years, owing to the potential benefits of this methodology in overcoming the limitations of active depth sensing systems. Moreover, due to the low cost and size of monocular cameras, researchers have focused their attention on monocular depth estimation (MDE), which consists in estimating a dense depth map from a single RGB video frame. State of the art MDE models typically rely on vision transformers (ViT) architectures that are highly deep and complex, making them unsuitable for fast inference on devices with hardware constraints. Purposely, in this paper, we address the problem of exploiting ViT in MDE on embedded devices. Those systems are usually characterized by limited memory capabilities and low-power CPU/GPU. We propose METER, a novel lightweight vision transformer architecture capable of achieving state of the art estimations and low latency inference performances on the considered embedded hardwares: NVIDIA Jetson TX1 and NVIDIA Jetson Nano. We provide a solution consisting of three alternative configurations of METER, a novel loss function to balance pixel estimation and reconstruction of image details, and a new data augmentation strategy to improve the overall final predictions. The proposed method outperforms previous lightweight works over the two benchmark datasets: the indoor NYU Depth v2 and the outdoor KITTI.