Unsupervised depth estimation from a single image is a very attractive technique with several implications in robotic, autonomous navigation, augmented reality and so on. This topic represents a very challenging task and the advent of deep learning enabled to tackle this problem with excellent results. However, these architectures are extremely deep and complex. Thus, real-time performance can be achieved only by leveraging power-hungry GPUs that do not allow to infer depth maps in application fields characterized by low-power constraints. To tackle this issue, in this paper we propose a novel architecture capable to quickly infer an accurate depth map on a CPU, even of an embedded system, using a pyramid of features extracted from a single input image. Similarly to state-of-the-art, we train our network in an unsupervised manner casting depth estimation as an image reconstruction problem. Extensive experimental results on the KITTI dataset show that compared to the top performing approach our network has similar accuracy but a much lower complexity (about 6% of parameters) enabling to infer a depth map for a KITTI image in about 1.7 s on the Raspberry Pi 3 and at more than 8 Hz on a standard CPU. Moreover, by trading accuracy for efficiency, our network allows to infer maps at about 2 Hz and 40 Hz respectively, still being more accurate than most state-of-the-art slower methods. To the best of our knowledge, it is the first method enabling such performance on CPUs paving the way for effective deployment of unsupervised monocular depth estimation even on embedded systems.
Given a classifier ensemble and a set of examples to be classified, many examples may be confidently and accurately classified after only a subset of the base models in the ensemble are evaluated. This can reduce both mean latency and CPU while maintaining the high accuracy of the original ensemble. To achieve such gains, we propose jointly optimizing a fixed evaluation order of the base models and early-stopping thresholds. Our proposed objective is a combinatorial optimization problem, but we provide a greedy algorithm that achieves a 4-approximation of the optimal solution for certain cases. For those cases, this is also the best achievable polynomial time approximation bound unless $P = NP$. Experiments on benchmark and real-world problems show that the proposed Quit When You Can (QWYC) algorithm can speed-up average evaluation time by $2$x--$4$x, and is around $1.5$x faster than prior work. QWYC's joint optimization of ordering and thresholds also performed better in experiments than various fixed orderings, including gradient boosted trees' ordering.
This paper proposes LPRNet - end-to-end method for Automatic License Plate Recognition without preliminary character segmentation. Our approach is inspired by recent breakthroughs in Deep Neural Networks, and works in real-time with recognition accuracy up to 95% for Chinese license plates: 3 ms/plate on nVIDIA GeForce GTX 1080 and 1.3 ms/plate on Intel Core i7-6700K CPU. LPRNet consists of the lightweight Convolutional Neural Network, so it can be trained in end-to-end way. To the best of our knowledge, LPRNet is the first real-time License Plate Recognition system that does not use RNNs. As a result, the LPRNet algorithm may be used to create embedded solutions for LPR that feature high level accuracy even on challenging Chinese license plates.
Neural architecture search (NAS) has been proposed to automatically tune deep neural networks, but existing search algorithms, e.g., NASNet, PNAS, usually suffer from expensive computational cost. Network morphism, which keeps the functionality of a neural network while changing its neural architecture, could be helpful for NAS by enabling more efficient training during the search. In this paper, we propose a novel framework enabling Bayesian optimization to guide the network morphism for efficient neural architecture search. The framework develops a neural network kernel and a tree-structured acquisition function optimization algorithm to efficiently explores the search space. Intensive experiments on real-world benchmark datasets have been done to demonstrate the superior performance of the developed framework over the state-of-the-art methods. Moreover, we build an open-source AutoML system based on our method, namely Auto-Keras. The system runs in parallel on CPU and GPU, with an adaptive search strategy for different GPU memory limits.
Medical image registration is an active research topic and forms a basis for many medical image analysis tasks. Although image registration is a rather general concept specialized methods are usually required to target a specific registration problem. The development and implementation of such methods has been tough so far as the gradient of the objective has to be computed. Also, its evaluation has to be performed preferably on a GPU for larger images and for more complex transformation models and regularization terms. This hinders researchers from rapid prototyping and poses hurdles to reproduce research results. There is a clear need for an environment which hides this complexity to put the modeling and the experimental exploration of registration methods into the foreground. With the "Autograd Image Registration Laboratory" (AIRLab), we introduce an open laboratory for image registration tasks, where the analytic gradients of the objective function are computed automatically and the device where the computations are performed, on a CPU or a GPU, is transparent. It is meant as a laboratory for researchers and developers enabling them to rapidly try out new ideas for registering images and to reproduce registration results which have already been published. AIRLab is implemented in Python using PyTorch as tensor and optimization library and SimpleITK for basic image IO. Therefore, it profits from recent advances made by the machine learning community concerning optimization and deep neural network models. The presented draft of this paper outlines AIRLab with first code snippets and performance analyses. A more exhaustive introduction will follow as a final version soon.
We introduce a principled approach for unsupervised structure learning of deep neural networks. We propose a new interpretation for depth and inter-layer connectivity where conditional independencies in the input distribution are encoded hierarchically in the network structure. Thus, the depth of the network is determined inherently. The proposed method casts the problem of neural network structure learning as a problem of Bayesian network structure learning. Then, instead of directly learning the discriminative structure, it learns a generative graph, constructs its stochastic inverse, and then constructs a discriminative graph. We prove that conditional-dependency relations among the latent variables in the generative graph are preserved in the class-conditional discriminative graph. We demonstrate on image classification benchmarks that the deepest layers (convolutional and dense) of common networks can be replaced by significantly smaller learned structures, while maintaining classification accuracy---state-of-the-art on tested benchmarks. Our structure learning algorithm requires a small computational cost and runs efficiently on a standard desktop CPU.
Neural networks have established as a generic and powerful means to approach challenging problems such as image classification, object detection or decision making. Their successful employment foots on an enormous demand of compute. The quantization of network parameters and the processed data has proven a valuable measure to reduce the challenges of network inference so effectively that the feasible scope of applications is expanded even into the embedded domain. This paper describes the making of a real-time object detection in a live video stream processed on an embedded all-programmable device. The presented case illustrates how the required processing is tamed and parallelized across both the CPU cores and the programmable logic and how the most suitable resources and powerful extensions, such as NEON vectorization, are leveraged for the individual processing steps. The crafted result is an extended Darknet framework implementing a fully integrated, end-to-end solution from video capture over object annotation to video output applying neural network inference at different quantization levels running at 16~frames per second on an embedded Zynq UltraScale+ (XCZU3EG) platform.
In many important applications -- such as search engines and relational database systems -- data is stored in the form of arrays of integers. Encoding and, most importantly, decoding of these arrays consumes considerable CPU time. Therefore, substantial effort has been made to reduce costs associated with compression and decompression. In particular, researchers have exploited the superscalar nature of modern processors and SIMD instructions. Nevertheless, we introduce a novel vectorized scheme called SIMD-BP128 that improves over previously proposed vectorized approaches. It is nearly twice as fast as the previously fastest schemes on desktop processors (varint-G8IU and PFOR). At the same time, SIMD-BP128 saves up to 2 bits per integer. For even better compression, we propose another new vectorized scheme (SIMD-FastPFOR) that has a compression ratio within 10% of a state-of-the-art scheme (Simple-8b) while being two times faster during decoding.
The change of two orders of magnitude in the 'new DCF' of NIST's SRE'10, relative to the 'old DCF' evaluation criterion, posed a difficult challenge for participants and evaluator alike. Initially, participants were at a loss as to how to calibrate their systems, while the evaluator underestimated the required number of evaluation trials. After the fact, it is now obvious that both calibration and evaluation require very large sets of trials. This poses the challenges of (i) how to decide what number of trials is enough, and (ii) how to process such large data sets with reasonable memory and CPU requirements. After SRE'10, at the BOSARIS Workshop, we built solutions to these problems into the freely available BOSARIS Toolkit. This paper explains the principles and algorithms behind this toolkit. The main contributions of the toolkit are: 1. The Normalized Bayes Error-Rate Plot, which analyses likelihood- ratio calibration over a wide range of DCF operating points. These plots also help in judging the adequacy of the sizes of calibration and evaluation databases. 2. Efficient algorithms to compute DCF and minDCF for large score files, over the range of operating points required by these plots. 3. A new score file format, which facilitates working with very large trial lists. 4. A faster logistic regression optimizer for fusion and calibration. 5. A principled way to define EER (equal error rate), which is of practical interest when the absolute error count is small.
Recently, machine learning algorithms have successfully entered large-scale real-world industrial applications (e.g. search engines and email spam filters). Here, the CPU cost during test time must be budgeted and accounted for. In this paper, we address the challenge of balancing the test-time cost and the classifier accuracy in a principled fashion. The test-time cost of a classifier is often dominated by the computation required for feature extraction-which can vary drastically across eatures. We decrease this extraction time by constructing a tree of classifiers, through which test inputs traverse along individual paths. Each path extracts different features and is optimized for a specific sub-partition of the input space. By only computing features for inputs that benefit from them the most, our cost sensitive tree of classifiers can match the high accuracies of the current state-of-the-art at a small fraction of the computational cost.
AdaBoost is an important algorithm in machine learning and is being widely used in object detection. AdaBoost works by iteratively selecting the best amongst weak classifiers, and then combines several weak classifiers to obtain a strong classifier. Even though AdaBoost has proven to be very effective, its learning execution time can be quite large depending upon the application e.g., in face detection, the learning time can be several days. Due to its increasing use in computer vision applications, the learning time needs to be drastically reduced so that an adaptive near real time object detection system can be incorporated. In this paper, we develop a hybrid parallel and distributed AdaBoost algorithm that exploits the multiple cores in a CPU via light weight threads, and also uses multiple machines via a web service software architecture to achieve high scalability. We present a novel hierarchical web services based distributed architecture and achieve nearly linear speedup up to the number of processors available to us. In comparison with the previously published work, which used a single level master-slave parallel and distributed implementation [1] and only achieved a speedup of 2.66 on four nodes, we achieve a speedup of 95.1 on 31 workstations each having a quad-core processor, resulting in a learning time of only 4.8 seconds per feature.
Superpixel algorithms aim to over-segment the image by grouping pixels that belong to the same object. Many state-of-the-art superpixel algorithms rely on minimizing objective functions to enforce color ho- mogeneity. The optimization is accomplished by sophis- ticated methods that progressively build the superpix- els, typically by adding cuts or growing superpixels. As a result, they are computationally too expensive for real-time applications. We introduce a new approach based on a simple hill-climbing optimization. Starting from an initial superpixel partitioning, it continuously refines the superpixels by modifying the boundaries. We define a robust and fast to evaluate energy function, based on enforcing color similarity between the bound- aries and the superpixel color histogram. In a series of experiments, we show that we achieve an excellent com- promise between accuracy and efficiency. We are able to achieve a performance comparable to the state-of- the-art, but in real-time on a single Intel i7 CPU at 2.8GHz.
HOL(y)Hammer is an online AI/ATP service for formal (computer-understandable) mathematics encoded in the HOL Light system. The service allows its users to upload and automatically process an arbitrary formal development (project) based on HOL Light, and to attack arbitrary conjectures that use the concepts defined in some of the uploaded projects. For that, the service uses several automated reasoning systems combined with several premise selection methods trained on all the project proofs. The projects that are readily available on the server for such query answering include the recent versions of the Flyspeck, Multivariate Analysis and Complex Analysis libraries. The service runs on a 48-CPU server, currently employing in parallel for each task 7 AI/ATP combinations and 4 decision procedures that contribute to its overall performance. The system is also available for local installation by interested users, who can customize it for their own proof development. An Emacs interface allowing parallel asynchronous queries to the service is also provided. The overall structure of the service is outlined, problems that arise and their solutions are discussed, and an initial account of using the system is given.
As a present to Mizar on its 40th anniversary, we develop an AI/ATP system that in 30 seconds of real time on a 14-CPU machine automatically proves 40% of the theorems in the latest official version of the Mizar Mathematical Library (MML). This is a considerable improvement over previous performance of large- theory AI/ATP methods measured on the whole MML. To achieve that, a large suite of AI/ATP methods is employed and further developed. We implement the most useful methods efficiently, to scale them to the 150000 formulas in MML. This reduces the training times over the corpus to 1-3 seconds, allowing a simple practical deployment of the methods in the online automated reasoning service for the Mizar users (MizAR).
We investigate fundamental decisions in the design of instruction set architectures for linear genetic programs that are used as both model systems in evolutionary biology and underlying solution representations in evolutionary computation. We subjected digital organisms with each tested architecture to seven different computational environments designed to present a range of evolutionary challenges. Our goal was to engineer a general purpose architecture that would be effective under a broad range of evolutionary conditions. We evaluated six different types of architectural features for the virtual CPUs: (1) genetic flexibility: we allowed digital organisms to more precisely modify the function of genetic instructions, (2) memory: we provided an increased number of registers in the virtual CPUs, (3) decoupled sensors and actuators: we separated input and output operations to enable greater control over data flow. We also tested a variety of methods to regulate expression: (4) explicit labels that allow programs to dynamically refer to specific genome positions, (5) position-relative search instructions, and (6) multiple new flow control instructions, including conditionals and jumps. Each of these features also adds complication to the instruction set and risks slowing evolution due to epistatic interactions. Two features (multiple argument specification and separated I/O) demonstrated substantial improvements int the majority of test environments. Some of the remaining tested modifications were detrimental, thought most exhibit no systematic effects on evolutionary potential, highlighting the robustness of digital evolution. Combined, these observations enhance our understanding of how instruction architecture impacts evolutionary potential, enabling the creation of architectures that support more rapid evolution of complex solutions to a broad range of challenges.
Principal component analysis (PCA) is a statistical technique commonly used in multivariate data analysis. However, PCA can be difficult to interpret and explain since the principal components (PCs) are linear combinations of the original variables. Sparse PCA (SPCA) aims to balance statistical fidelity and interpretability by approximating sparse PCs whose projections capture the maximal variance of original data. In this paper we present an efficient and paralleled method of SPCA using graphics processing units (GPUs), which can process large blocks of data in parallel. Specifically, we construct parallel implementations of the four optimization formulations of the generalized power method of SPCA (GP-SPCA), one of the most efficient and effective SPCA approaches, on a GPU. The parallel GPU implementation of GP-SPCA (using CUBLAS) is up to eleven times faster than the corresponding CPU implementation (using CBLAS), and up to 107 times faster than a MatLab implementation. Extensive comparative experiments in several real-world datasets confirm that SPCA offers a practical advantage.
We present a novel, realtime algorithm to compute the trajectory of each pedestrian in moderately dense crowd scenes. Our formulation is based on an adaptive particle filtering scheme that uses a multi-agent motion model based on velocity-obstacles, and takes into account local interactions as well as physical and personal constraints of each pedestrian. Our method dynamically changes the number of particles allocated to each pedestrian based on different confidence metrics. Additionally, we use a new high-definition crowd video dataset, which is used to evaluate the performance of different pedestrian tracking algorithms. This dataset consists of videos of indoor and outdoor scenes, recorded at different locations with 30-80 pedestrians. We highlight the performance benefits of our algorithm over prior techniques using this dataset. In practice, our algorithm can compute trajectories of tens of pedestrians on a multi-core desktop CPU at interactive rates (27-30 frames per second). To the best of our knowledge, our approach is 4-5 times faster than prior methods, which provide similar accuracy.
The amount of information in the form of features and variables avail- able to machine learning algorithms is ever increasing. This can lead to classifiers that are prone to overfitting in high dimensions, high di- mensional models do not lend themselves to interpretable results, and the CPU and memory resources necessary to run on high-dimensional datasets severly limit the applications of the approaches. Variable and feature selection aim to remedy this by finding a subset of features that in some way captures the information provided best. In this paper we present the general methodology and highlight some specific approaches.
Computational Grids are a new trend in distributed computing systems. They allow the sharing of geographically distributed resources in an efficient way, extending the boundaries of what we perceive as distributed computing. Various sciences can benefit from the use of grids to solve CPU-intensive problems, creating potential benefits to the entire society. Job scheduling is an integrated part of parallel and distributed computing. It allows selecting correct match of resource for a particular job and thus increases the job throughput and utilization of resources. Job should be scheduled in an automatic way to make the system more reliable, accessible and less sensitive to subsystem failures. This paper provides a survey on various heuristic algorithms, used for scheduling in grid.
In this paper, we present an OpenCL-based heterogeneous implementation of a computer vision algorithm -- image inpainting-based object removal algorithm -- on mobile devices. To take advantage of the computation power of the mobile processor, the algorithm workflow is partitioned between the CPU and the GPU based on the profiling results on mobile devices, so that the computationally-intensive kernels are accelerated by the mobile GPGPU (general-purpose computing using graphics processing units). By exploring the implementation trade-offs and utilizing the proposed optimization strategies at different levels including algorithm optimization, parallelism optimization, and memory access optimization, we significantly speed up the algorithm with the CPU-GPU heterogeneous implementation, while preserving the quality of the output images. Experimental results show that heterogeneous computing based on GPGPU co-processing can significantly speed up the computer vision algorithms and makes them practical on real-world mobile devices.