Networks are widely used in modern communication, social, and technological systems, but their structure and dynamic processes can be difficult to understand through static visualizations alone. This thesis presents the design and implementation of a set of handheld wireless controllers that function as physical nodes in a network. By moving the controllers, users can create and modify network topologies interactively, making abstract network concepts more tangible for educational use.The system consists of ESP32-based microcontroller nodes that communicate using the ESP-NOW protocol. Each node determines its connections to other nodes using received signal strength indicator measurements, which are smoothed and interpreted through a double-boundary method to improve connection stability. A coordinator node communicates with a host computer, where a Python-based user interface visualizes the network topology and node states in real time. In addition to forming networks, the controllers can be used to run a dynamic process on the network: the Non-Consensus Opinion model. In this model, each node holds one of two opinions and updates its state based on the majority opinion in its local neighbourhood. The implementation was tested on several network structures, including line, circle, and island configurations. The experimental results show that the physical controller network can reproduce the expected behaviour of the opinion model while providing both local LED feedback and on-screen visualization. The project demonstrates that handheld wireless controllers can provide an interactive and intuitive platform for teaching network topology and dynamic processes on networks. Future improvements may further increase communication efficiency, positioning accuracy, and scalability. ... Read more

Disinfection of seeds is a method used by the seed industry to remove pathogens. However, conventional methods have their shortcomings in terms of energy efficiency and yield. This thesis focuses on the design of a fluidized bed setup with sensors for the purpose of DBD plasma disinfection.
Two configurations were investigated, namely a spouted bed and a bubbling fluidized bed, where the bubbling fluidized bed produced the best results. Additionally, the fluidized bed was equipped with motor speed control, and with sensors for the measurement of ozone concentration, temperature, humidity, air speed, and ion density. These sensors were integrated in an easy readout system using displays, making the system ready for testing of disinfection performance.
... Read more

As a novel alternative for conventional seed disinfection methods, a new design has been proposed in this report using a surface dielectric barrier discharge (SDBD) fractal electrode. The discharge mechanism for this electrode is a diffuse microdischarge under AC or short-pulsed DC mode operation. In this way, cold plasma could be generated that is applicable for seed disinfection. Furthermore, the electrodes were designed to be part of a proof of concept fluidized bed reactor with a reactor size of 10×20×20[𝑐𝑚] for disinfecting cabbage seeds with a diameter of 2𝑚𝑚. Because of this application, the efficacy of seed decontamination using plasma with its generated reactive agents was discussed. The used gas mixture in which the electrodes created plasma was ambient air without increased humidity. This means that the main reactive agents for sterilisation are reactive oxygen species (RON) like ozone (O3) and reactive nitrogen species (RNS). The electrical and physical parameters required to make cold-plasma were investigated to come up with a proper design for the electrode. From this theoretical analysis, five different initial designs emerged. The analysed designs include a wire-to-wire, wire-to-sheet, multi-hollow DBD, fractal SDBD and a coplanar DBD fractal electrode. All electrode designs were made based on the state-of-the-art dielectric barrier discharge principle. Moreover, in the design consideration, different materials for the conductor and dielectric were discussed, mainly based on electrical properties, plasma generation and manufacturing possibilities were considered. Based on previously set trade-off requirements, together with the results of measured power and turn on voltage of the plasma electrodes, the best designs tested design for seed disinfection are the double-sided 5𝑡ℎ order Hilbert fractal with a 1.6 𝑚𝑚 barrier and the single-sided 5𝑡ℎ order Hilbert fractal with a 0.8 𝑚𝑚 barrier. ... Read more

Negotiation Support Systems (NSSs) can provide help based on the preference setting (domain, issue weights, issue ranking, strategies, etc.) of the users of the systems. However, sometimes the users of the systems might make mistakes in the preference setting. With wrong preferences, the NSSs might provide suggestions that conflict with the users' desires. This thesis focuses on designing a mechanism that could detect the mismatches in issue weights of the users of NSSs and provide recommendations on updating the issue weights with a scoring method. This mechanism also has the ability to provide different recommendations based on the users' self-confidence level. A case study is done with six simulation experiments to test the mechanism. The results show that during the negotiation, the users of the NSS (PN) can have a chance to change the issue weights when mismatches occurred and the NSS (PN) can provide better suggestions after the issue weights are updated. The limitations of the mechanism are concluded at the end together with the future thoughts ... Read more

This report details the design and development of an agar/NaCl gel-like tissue phantom mimicking the electrical properties of wet human skin. The skin phantom provides a reliable, reproducible testing ground for dry-contact polydimethylsiloxane (CNT/PDMS) electrodes, with the aim of recording electroencephalograms (EEGs) and stimulating brain activity in a controlled environment. These electrodes are being designed for the development of an in-ear brain-computer interface (BCI).
The electrical properties of biological tissue are referred to as the conductivity σ and permittivity ε and denote the ability for a material to conduct and trap electric charge respectively. These properties are frequency dependent and particularly for EEGs, a frequency range of 1-1000 Hz is of interest (with some added leeway). Wet skin hereby has a conductivity of around 0.1 Siemens to 0.2 Siemens in the 1-1000 Hz frequency range whereas the permittivity ranges from 5.7 * 10^5 to 5.2 * 10^5. Different agar and agar/NaCl solutions are created to try and obtain solutions with the mentioned electric properties. Specifically, NaCl is added to improve the conductivity and obtain a non-linear frequency response similar to that of human skin. The electrical properties of the phantoms were verified/measured using the parallel plate method. This method is essentially sandwiching a material under test (MUT) (in this case the fabricated gel-like agar and agar/NaCl solutions) between two conducting plates. This method is most suited for measurements in the lower frequency spectrum.
The skin phantom consisting of 3.04 mass fraction weight (wt.%) agar and 0.539 wt.% NaCl shows the closest similarity to the conductivity of wet skin. Namely, a conductivity of ~ 0.1 Siemens to 0.45 Siemens in the frequency range of 1-1000 Hz. A decrease of 0.250 wt.% NaCl will most likely achieve the desired conductivity response of 0.1 Siemens to 0.2 Siemens in the frequency range of 1-1000 Hz. The skin phantom consisting of 3.00 wt.% agar and 1.02 wt.% NaCl showed the permittivity closest to that of wet skin, but might have been a noisy outlier. Its permittivity ranges from 10 * 10^6 and 7.5 * 10^6. This is still a large error margin from the desired 5.7 * 10^5 to 5.2 * 10^5. Additional fillers like glycine or Al powder need to be added to the solutions to obtain a permittivity close to that of wet human skin. Multi-day and difference in applied pressure measurements are performed to check the sensitivity and reproducibility of the phantoms. Applied pressure hereby has little to no influence whereas a longer life-span of the fabricated phantom shows a drastic decrease of the electrical properties of the phantoms after day 1. The changes then seem to settle. Worth mentioning is that the change is only drastic when the solution has a high conductivity. This is generally not the case for solutions with conductivities close to wet skin. ... Read more

Global Navigation Satellite Systems (GNSSs) have become a critical part of the infrastructure of modern society. Radio interference can introduce position or timing errors in systems that use GNSS or, in a worst-case scenario, block the reception of GNSS signals in full. Part of this critical infrastructure is, among others, power plants, banks, and transport. Interference of GNSS signals could originate from nature, such as solar activity or ionospheric effects. Other interference could originate from unintentional sources (e.g., radio signals from a malfunctioning radio tower) or intentional sources such as a jamming or spoofing device. The latter is what this thesis will focus on. This thesis consists of two parts. The first part is about jamming and elaborates on the impact of seven different forms of jamming on two types of GNSS receivers, a time-worn receiver and a cuttingedge receiver. The cutting-edge receiver has as option to turn on Interference Mitigation (IM). The performance of both receivers is the roughly the same in case the IM is turned off on the cutting-edge receiver. However, when the IM is turned on the cutting-edge receiver clearly is more resillient to the jamming signals. In the second part of the thesis various types of spoofing are discussed. Due to time and hardware restrictions it was not possible to perform synchronous spoofing, which is an advanced form of spoofing. Instead, various concepts are discussed that describe how synchronous spoofing could be achieved. ... Read more

Resting-state fMRI (rs-fMRI) has become an important imaging modality and is commonly used to study intrinsic brain networks. These networks can be obtained by decomposing rs-fMRI data into components, using independent component analysis (ICA). Recently, these ICA components have been used as inputs for neural networks to learn complex relations between the intrinsic networks of the brain and mental disorders or demographic variables. Instead of training a non-linear classifier on these linearly decomposed components, this work asks whether unsupervised representation learning can lead to linearly separable representations for multiple downstream tasks.

We propose to apply non-linear representation learning to voxelwise rs-fMRI data. Learning the non-linear representations is done using two versions of a variational autoencoder (VAE). The first version is a vanilla VAE with 3D residual blocks in both its encoder and decoder. The second version is based on the identifiable VAE and uses a time-dependent prior. The models train to reconstruct the original input data from latent variables it infers. Three predictive models then evaluate the predictive power of the latent variables on an age regression, a sex classification, and a schizophrenia classification task. Each of the predictive models performs predictions for each of the three tasks. The predictive models are a support vector machine (SVM), a k-nearest neighbor (k-NN) model, and a long short-term memory (LSTM) neural network.

We show that our method performs exceptionally well on the age regression and sex classification tasks without any supervision. These results imply that VAEs can model predictive variations in their latent spaces for demographic variables. The models, however, do not do well on the schizophrenia classification task, even when the models are pretrained. Despite the lower performance on the schizophrenia classification task, the overall results are encouraging and pave the way for future work on voxelwise representation learning. ... Read more

With modern multi-function radars becoming more flexible, handling the limited amount of resources of these radars becomes increasingly important. In this thesis the radar resource management (RRM) problem in a multi-target tracking scenario is considered. Partially observable Markov decision processes (POMDPs) are used to describe each tracking task. By comparing the future effect of radar actions using model predictive control (MPC), the POMDPs are solved in a non-myopic way. The model predictive control problem can be decoupled into sub-problems using Lagrangian Relaxation to reduce the computational complexity of the solution method. An algorithm based on golden section search is employed to find the Lagrange multiplier. An interacting multiple model filter is used to allow the method to be effective in RRM problems involving the tracking of targets performing a broad number of maneuvers.
The novel approach is compared to an existing solution method based on policy rollout and Monte Carlo sampling. Through simulations of dynamic multi-target tracking scenarios in which the cost and computational complexity of different approaches are compared, it was shown that the computational complexity is greatly reduced while the resulting resource allocation results remain similar. ... Read more

In the past years, small Earth Observation (EO) satellites have become increasingly capable of taking high-resolution images at high sample rates. These images contain valuable information for different sectors, such as the agricultural and military sector. Furthermore they can contain important information about the climate and climate change. Sending these images to earth requires a large amount of down-link bandwidth. This results in heavy, large power modules and communication modules, resulting in larger, more expensive (in terms of launch cost as well as in terms of production cost) satellites. This phenomenon already results in satellites not sending all information they gather, with examples of being able to send 2 minutes worth of data per orbit (approx. 90 minutes) not being out of the ordinary. As more and more satellites are transmitting data towards earth the communication is also expected to become even more power-intensive (or even more limited), since the (theoretically) available bandwidth per satellite is reduced. Therefore a shift towards a different approach is necessary. Smarter ways to get the relevant information to earth have to be developed. In contrast with the "common knowledge" that is often applied in the field of object detection, using the highest possible image quality does not transfer to the best trained network when gathering a dataset of satellite images, since the main constraint is the bandwidth available for transmitting images, where "normally" the largest constraint is the amount of man-hours spent on annotation. Compression of training images with JPEG-XR quality level 2 during the gathering of training images results in a better "bandwidth-efficiency", in the dataset used for this research at least up to 30000 images. It was also found that when more bandwidth is available and thus more images can be added to the training set, the optimal amount of compression tends to decrease. This results also lies in line with the result that the "final accuracy" (the predicted accuracy of a model trained on an infinite amount of images) of the models tend to improve with better image quality. From this it can be concluded that for the optimal approach the training images should be compressed as far as possible at the start of training, to then decrease the amount of compression as the mission progresses and more cumulative bandwidth is available. ... Read more

How do automatic object detector outputs align with what humans consider good object detection? Our study is based on the responses of 70 participants for a survey. The participants are presented with images having bound- ing box predictions, their task is to choose images which according to them have an acceptable or a good detection. The results show a correlation between the size of the object and the evaluation metric IoU (Intersection over Union), with the size of the bounding box. Furthermore, the data indicates that the kind of box they prefer most for a detection output, is also the most accepted detection by them. Additionally, the results suggest that based on the symmetry of the object, position of the bounding box may or may not play a role for considering a detection valid. Our study investigates through human subjective choices if the traditional threshold value of IoU for evaluation, and tight bounding box outputs are always the best outputs in object detection techniques. ... Read more

Decreasing injuries in football is a topic of interest for the KNVB and KNHB. To reach this goal, the use of smart sensor pants is researched. The data will be used to develop models for finding injury risk factors that are related to movements. Currently, the system is capable of reading out the IMUs and storing the data on an SD-card, for post-analysis of the data. As next step, the communication link should be implemented, for real time feedback to the football players. The design of an efficient communication system in terms of power, size, and reliability, is quite a big task. By looking at similar research to body-worn devices, it is noted, that most devices deal with a 10 times lower data rate than the current device. On top of that, it is noted that the absorption of the body is causing problems on the link reliability. Based on these observations, it is decided to focus the research of this Thesis on the reduction of the data load, and the optimization of the transmission part of the smart sensor pants. The Thesis is split into two parts, the first part will show the use of lossless data compression. The second part will start with showing the benefit of using a patch antenna over a dipole antenna and continues by showing the benefit of using a dual antenna configuration over a single antenna configuration. To be more specific, the first part of this Thesis starts by comparing different lossless data compression algorithms, from which the FELACS algorithm is chosen as most suited. This algorithm is then implemented on the current hardware and tested on data from the smart sensor pants in a realistic football scenario. These results show an average compression ratio of 43 %-45 % in the most intensive 5-minutes of a football game, with a minimum of 38 % in an interval of 10 s. To improve the compression algorithm, an adjustment to the FELACS algorithm is proposed. This adjustment is theoretically tested and shown to outperform the FELACS algorithm with a higher compression ratio. In the second part, the use of a dual antenna configuration is discussed, whereby the use of a patch antenna is compared to a dipole antenna. It will be shown, that a dual antenna configuration can significantly improve the signal strength around the player, resulting in an almost isotropic radiation pattern, using pattern diversity. On top of that, this form of pattern diversity is observed to increase the reliability of the link, using switched combining. Moreover, it will be shown that a patch antenna will be more suited for this application, due to the higher gain in the front, and the robustness against interference when placed close to a conducting material. In summary, the two main contributions of this Thesis, are the reductions in data load, and the testing and verification of the dual patch antenna configuration. These contributions provide the basis for the communication part of the smart sensor pants. ... Read more

2.5D shape display is a recent idea in the market that emerged as a platform of interaction between a computer and human. 2.5D shape display is essentially a grid-like matrix consisting of actuators and pins moving up and down in vertical motion to create pseudo-3D images. Focused as a visual display in some applications and as a medium of input or output in others, this technology holds a lot of potentials to be explored in present and future applications since it is a unique type of hardware that can actually show images in a real world. Recently, many such platforms have been created by the use of different hardware solutions and have found applications in gaming interface, physical telepresence, dynamic objects etc. Some research areas that can still be explored are applications for the blind, building and object modelling etc. With advances in applications, there is a need for hardware with higher resolution at a lower cost. Also, being an interact-able display, safety and comfort of the user is of utmost importance. This has been seen to be lacking in the existing projects. In this thesis, we focus on designing a prototype for building a safe display with minimal cost. We then go on to understand safety and comfort for the chosen display prototype and design some safe actuation algorithms. These are later evaluated using a combination of an experimental survey and simulation to find and propose a good solution for safe 2.5D shape displays. ... Read more

Software bugs in many different variants can potentially leak sensitive data to an attacker. Implementing a separation mechanism for security domains can prevent incorrect or malicious code to leak sensitive data from one security domain to another. This work presents a separation mechanism based on labeling security domains with a label in tagged memory, at word-level granularity, called color labeling.

Utilizing a tagged architecture based on the RISC-V architecture, color labeling assigns colors (denoting a security domain) to individual memory words, cache lines, registers and peripherals. Using a simple set of hardware enforced policies, data protection is ensured. Control flow integrity is maintained with the
help of additional tag bits that denote code and valid jump addresses. New instructions have been added for functions that handle data residing in multiple security domains.

Software support is implemented in the Rust compiler. The compiler is enhanced with macros to support the coloring concept via source level annotations. Incorrect use of labels is reported during compilation. An external tool is used to generate tag information and generate a security report with information on
variable coloring and special function use and construction. Using the external tool keeps the changes to the compiler minimal, thereby reducing the maintenance burden and the required trust in the compiler as well. The report can be used in a security audit.

The concept is implemented on an instruction set architecture simulator. The toolchain modifications and the concept itself have been tested on this simulator. Testing showed the concept can prevent cross-security domain information leaks under several common attack patterns. The overhead due to the execution of additional instructions in the executable code depends on the actual code. Tests with the typical target application OpenVPN-NL showed a less than 5% increase in instruction count for the most commonly called functions.

By designing or redesigning software specifically for color labeling, this overhead can possibly be further reduced. Further testing, specifically on an actual hardware implementation is recommended.

Due to timing constraints, the concept has not been implemented in hardware. However, the hardware performance costs are estimated to be negligible. The area requirements are substantial: implementing the concept in the RISC-V softcore requires double the external memory capacity and FPGA resource utilization is estimated to require 14% more ALMs and 74% more internal memory blocks. ... Read more

Air travel has become considerably safe in the last decades, yet we see some fatal accidents. Some recent incidents reveal that there are issues in the measurement of wind speed and angle of attack (AoA). This paper presents a novel system for concurrently sensing the wind speed and AoA of an aircraft. We present the design of a wireless sensor, called Hermes, that simultaneously enables sensing as well as piezoelectric energy harvesting, making it self-powered and batteryless. Hermes comprises of a set of piezoelectric films which flutter due to incoming wind speed, and the characteristic of this aeroelastic flutter is utilized for determining the wind speed and AoA of the incoming airflow. The design of Hermes is such that sensing performance and energy harvesting capability are simultaneously maximized, with- out the need for trading off.

Hermes, a small form factor electronic module along with piezoelectric films and a 3D mount, is fabricated, tested in a wind tunnel, and in a real aircraft fuselage for its communication performance. Hermes harvests an average power of 440 휇W power. Over a wide range of AoA of −10 to 30 degrees, the estimation of the wind speed is within 0.2 m/s error with 90% probability and AoA error is within 1.2 degree with 90% probability. Hermes can be used in light aircraft and long endurance UAVs as is. It can also be used in several other applications, such as windmills. Hermes is expected to open up new avenues for interdisciplinary research for aerospace applications.
... Read more

We present a new MAC protocol for networks of devices. We specifically target certain applications. To cater for this setting, we introduce a new concept. This concept is instrumental to improve performance in these network scenarios. We build a proof of concept implementation of the new mac protocol. We specifically present a protocol model and formulate the mathematical framework to study the performance of a network of devices following the new MAC. Comparing the new MAC against main contention-based MAC protocols, we show that it outperforms pure Aloha and 1-CSMA. As traffic increases, the energy savings achieved by the new MAC against CSMA protocols increase, showcasing its capability to scale. ... Read more

This thesis report focuses on possible methods of digital implementation of motional feedback in a bass loudspeaker. In this thesis, the control system will be created for a monopole and a dipole speaker specifically. It does so by sketching the outlines of such a system and examining possible designs for the components required to suppress distortion which is mainly created by the speaker. This thesis shows that the digital implementation indeed allows filtering with very little latency but there are limitations in accuracy due to the conversion from the analogue to the digital domain. It is shown in this thesis that a partly analogue and partly digital system is desired for better error correction. A desired input-output gain of 30=29.54dB is achieved. The controller that is suggested is a controller based on the transfer function of the speaker. By transforming the input signal with the inverse of this open-loop transfer, the closed-loop transfer will be flattened. The instrumentation amplifier used to subtract the feedback signal from the input signal is designed such that the influence of the quantization noise is minimized. Loop-gain is added by the instrumentation amplifier, the controller and voltage-to-current amplifier. By creating loop-gain as much as possible using the digital implementation, noise is suppressed from 30dB up to 90dB within the range of frequencies of interest. The designed components in the motional feedback system allow a loop-gain up to 70dB and an input to output voltage gain of 30dB before instability occurs. ... Read more

Imitation Learning is a technique that enables programming the behavior of agents through demonstration, as opposed to manually engineering behavior. However, Imitation Learning methods require demonstration data (in the form of state-action labels) and in many scenarios, the demonstrations can be expensive to obtain or too complex for a demonstrator to execute. This lack or sub-optimality of demonstrations limits the applicability and performance of many Imitation Learning methods.

Advancements in Interactive Imitation Learning techniques however, have made it easier for demonstrators to train agents and improve their performance. These techniques involve demonstrators interacting with and guiding the agent as it performs the requisite task. This guidance is typically in the form of corrections or feedback on the current actions being executed by the agent.

In this thesis, a novel Interactive Learning technique is proposed that uses human corrective feedback in state-space to train and improve agent behavior. This technique is beneficial since providing guidance to the agent in terms of `changing its state' is often easier or more intuitive for the human demonstrator (as opposed to changing the actions being executed). For instance, in manipulation tasks using a robotic arm, it is easier for the demonstrator to provide state information such as the Cartesian position of the end-effector rather than low-level action information such as joint angles. Keeping such scenarios in mind, we propose our method titled: Teaching Imitative Policies in State-space (TIPS).

We evaluate the performance of TIPS for various control tasks as part of the OpenAI Gym toolkit as well as for a manipulation task using a KUKA LBR iiwa robotic arm. We show that through continuous improvement via feedback, agents trained using TIPS outperform the demonstrator and in-turn outperform conventional Imitation Learning agents. ... Read more

Long Range Wide Area Networks (LoRaWAN) offer easy deployment, robustness against interference, and operational longevity to energy constrained IoTdevices which communicate in a best-e_ort fashion in extended ranges. However, the simple (ALOHA-like) design of the MAC layer leads to packet collisions in dense LoRaWAN deployments with high traffic loads. To achieve scalability above a few hundreds of devices, time division is not an option, since LoRaWAN is asynchronous regarding communication. Further, feedback mechanisms are discouraged due to duty cycle limitations. In this document, we propose Spreading Factor MAC (SFMAC); a distributed and energy efficient MAC protocol for LoRaWAN, wherein {for the _rst time to the best of our knowledge{ high-SF channels are dedicated strictly to Channel Sensing (CS), while low-SF channels are focused on data-transmission. The Capture E_ect (CE) phenomena that is manifested in the PHY layer is extensively evaluated on-_eld and embodied in the SFMAC operating principle. The dedicated high-SF sensing allows e_ective revealing of hidden devices' transmissions without a_ecting the low-SF tra_c. We showcase the impact of SFMAC in scalability by designing a realistic implementation of the mechanism in ns-3. We report a x2.08 improvement in channel utilization and x2 goodput compared to LoRaWAN, without substantially increasing complexity. ... Read more

UnmannedAerial Vehicles(UAVs) have multi-domain applications and fixed-wing UAVs are a widely used class. There is ongoing research on topics in view to optimize the control and guidance of UAVs. This work explores the design, implementation and Software-in-the-Loop validation of an autopilot using adaptive guidance laws with emphasis on formation control of multiple fixed-wing UAVs. The work is done on Raspberry Pis in C++ which can be interfaced to standard autopilots as companion computers. The work splits a mission given by the user into primitive missions and uses an adaptive vector field approach for following it. For formation control, the work implements a discretized version of the Model Reference Adaptive Control synchronisation laws for multi-agent systems. Simulations are done in a distributed setting with a server program designed for the purpose. The server program handles the user inputs and configurations of the UAVs. ... Read more

In the past few years, convolutional neural networks (CNNs) have been widely utilized and shown state-of-the-art performances on computer vision tasks. However, CNN based approaches usually require a large amount of storage, run-time memory, as well as computation power in both training and inference time, which are usually used on GPU based machines to ensure the speed for inferences. But they are usually insufficient to be deployed on low-power applications. Although many approaches were proposed to compress and accelerate the CNN models, most of them were only evaluated on relatively simple problems (e.g. image classification), which only support limited real-world applications. Especially, among those methods, binary quantization can achieve very high model compression, but only a few works have been observed to utilize it on more complex tasks. Therefore, the exploration and evaluations of applying binary quantization on more complex tasks like object detection are worthwhile, which can be used in much more applications like autonomous driving and face detection. In this project, we apply and evaluate two different binary quantization approaches, named ABC-Net and PA-Net on object detection tasks. Also, we specify the exact implementation details for the binary convolutional operations in this project. As a result, we can achieve maximally 6.1× (around 16% of the full-precision model) compression, and minimal 2.5% accuracy reduction for weight quantization. The weight quantized models were able to outperform some existing real-time detectors in terms of both accuracy and storage size. Although large accuracy reduction was observed for input quantization, the quantized model could still maintain an acceptable accuracy compared to existing real-time object detectors. ... Read more