We provide an open-source C++ library for real-time metric-semantic visual-inertial Simultaneous Localization And Mapping (SLAM). The library goes beyond existing visual and visual-inertial SLAM libraries (e.g., ORB-SLAM, VINS- Mono, OKVIS, ROVIO) by enabling mesh reconstruction and semantic labeling in 3D. Kimera is designed with modularity in mind and has four key components: a visual-inertial odometry (VIO) module for fast and accurate state estimation, a robust pose graph optimizer for global trajectory estimation, a lightweight 3D mesher module for fast mesh reconstruction, and a dense 3D metric-semantic reconstruction module. The modules can be run in isolation or in combination, hence Kimera can easily fall back to a state-of-the-art VIO or a full SLAM system. Kimera runs in real-time on a CPU and produces a 3D metric-semantic mesh from semantically labeled images, which can be obtained by modern deep learning methods. We hope that the flexibility, computational efficiency, robustness, and accuracy afforded by Kimera will build a solid basis for future metric-semantic SLAM and perception research, and will allow researchers across multiple areas (e.g., VIO, SLAM, 3D reconstruction, segmentation) to benchmark and prototype their own efforts without having to start from scratch.
Data loading can dominate deep neural network training time on large-scale systems. We present a comprehensive study on accelerating data loading performance in large-scale distributed training. We first identify performance and scalability issues in current data loading implementations. We then propose optimizations that utilize CPU resources to the data loader design. We use an analytical model to characterize the impact of data loading on the overall training time and establish the performance trend as we scale up distributed training. Our model suggests that I/O rate limits the scalability of distributed training, which inspires us to design a locality-aware data loading method. By utilizing software caches, our method can drastically reduce the data loading communication volume in comparison with the original data loading implementation. Finally, we evaluate the proposed optimizations with various experiments. We achieved more than 30x speedup in data loading using 256 nodes with 1,024 learners.
Machine learning is yielding unprecedented interest in research and industry, due to recent success in many applied contexts such as image classification and object recognition. However, the deployment of these systems requires huge computing capabilities, thus making them unsuitable for embedded systems. To deal with this limitation, many researchers are investigating brain-inspired computing, which would be a perfect alternative to the conventional Von Neumann architecture based computers (CPU/GPU) that meet the requirements for computing performance, but not for energy-efficiency. Therefore, neuromorphic hardware circuits that are adaptable for both parallel and distributed computations need to be designed. In this paper, we focus on Spiking Neural Networks (SNNs) with a comprehensive study of information coding methods and hardware exploration. In this context, we propose a framework for neuromorphic hardware design space exploration, which allows to define a suitable architecture based on application-specific constraints and starting from a wide variety of possible architectural choices. For this framework, we have developed a behavioral level simulator for neuromorphic hardware architectural exploration named NAXT. Moreover, we propose modified versions of the standard Rate Coding technique to make trade-offs with the Time Coding paradigm, which is characterized by the low number of spikes propagating in the network. Thus, we are able to reduce the number of spikes while keeping the same neuron's model, which results in an SNN with fewer events to process. By doing so, we seek to reduce the amount of power consumed by the hardware. Furthermore, we present three neuromorphic hardware architectures in order to quantitatively study the implementation of SNNs. These architectures are derived from a novel funnel-like Design Space Exploration framework for neuromorphic hardware.
We present an overview of SURREAL-System, a reproducible, flexible, and scalable framework for distributed reinforcement learning (RL). The framework consists of a stack of four layers: Provisioner, Orchestrator, Protocol, and Algorithms. The Provisioner abstracts away the machine hardware and node pools across different cloud providers. The Orchestrator provides a unified interface for scheduling and deploying distributed algorithms by high-level description, which is capable of deploying to a wide range of hardware from a personal laptop to full-fledged cloud clusters. The Protocol provides network communication primitives optimized for RL. Finally, the SURREAL algorithms, such as Proximal Policy Optimization (PPO) and Evolution Strategies (ES), can easily scale to 1000s of CPU cores and 100s of GPUs. The learning performances of our distributed algorithms establish new state-of-the-art on OpenAI Gym and Robotics Suites tasks.
The outstanding computational efficiency of discriminative correlation filter (DCF) fades away with various complicated improvements. Previous appearances are also gradually forgotten due to the exponential decay of historical views in traditional appearance updating scheme of DCF framework, reducing the model's robustness. In this work, a novel tracker based on DCF framework is proposed to augment memory of previously appeared views while running at real-time speed. Several historical views and the current view are simultaneously introduced in training to allow the tracker to adapt to new appearances as well as memorize previous ones. A novel rapid compressed context learning is proposed to increase the discriminative ability of the filter efficiently. Substantial experiments on UAVDT and UAV123 datasets have validated that the proposed tracker performs competitively against other 26 top DCF and deep-based trackers with over 40 FPS on CPU.
Object detection has made impressive progress in recent years with the help of deep learning. However, state-of-the-art algorithms are both computation and memory intensive. Though many lightweight networks are developed for a trade-off between accuracy and efficiency, it is still a challenge to make it practical on an embedded device. In this paper, we present a system-level solution for efficient object detection on a heterogeneous embedded device. The detection network is quantized to low bits and allows efficient implementation with shift operators. In order to make the most of the benefits of low-bit quantization, we design a dedicated accelerator with programmable logic. Inside the accelerator, a hybrid dataflow is exploited according to the heterogeneous property of different convolutional layers. We adopt a straightforward but resource-friendly column-prior tiling strategy to map the computation-intensive convolutional layers to the accelerator that can support arbitrary feature size. Other operations can be performed on the low-power CPU cores, and the entire system is executed in a pipelined manner. As a case study, we evaluate our object detection system on a real-world surveillance video with input size of 512x512, and it turns out that the system can achieve an inference speed of 18 fps at the cost of 6.9W (with display) with an mAP of 66.4 verified on the PASCAL VOC 2012 dataset.
The effective utilization at scale of complex machine learning (ML) techniques for HEP use cases poses several technological challenges, most importantly on the actual implementation of dedicated end-to-end data pipelines. A solution to these challenges is presented, which allows training neural network classifiers using solutions from the Big Data and data science ecosystems, integrated with tools, software, and platforms common in the HEP environment. In particular, Apache Spark is exploited for data preparation and feature engineering, running the corresponding (Python) code interactively on Jupyter notebooks. Key integrations and libraries that make Spark capable of ingesting data stored using ROOT format and accessed via the XRootD protocol, are described and discussed. Training of the neural network models, defined using the Keras API, is performed in a distributed fashion on Spark clusters by using BigDL with Analytics Zoo and also by using TensorFlow, notably for distributed training on CPU and GPU resourcess. The implementation and the results of the distributed training are described in detail in this work.
Image cropping aims to improve the composition as well as aesthetic quality of an image by removing extraneous content from it. Most of the existing image cropping databases provide only one or several human-annotated bounding boxes as the groundtruths, which can hardly reflect the non-uniqueness and flexibility of image cropping in practice. The employed evaluation metrics such as intersection-over-union cannot reliably reflect the real performance of a cropping model, either. This work revisits the problem of image cropping, and presents a grid anchor based formulation by considering the special properties and requirements (e.g., local redundancy, content preservation, aspect ratio) of image cropping. Our formulation reduces the searching space of candidate crops from millions to no more than ninety. Consequently, a grid anchor based cropping benchmark is constructed, where all crops of each image are annotated and more reliable evaluation metrics are defined. To meet the practical demands of robust performance and high efficiency, we also design an effective and lightweight cropping model. By simultaneously considering the region of interest and region of discard, and leveraging multi-scale information, our model can robustly output visually pleasing crops for images of different scenes. With less than 2.5M parameters, our model runs at a speed of 200 FPS on one single GTX 1080Ti GPU and 12 FPS on one i7-6800K CPU. The code is available at: \url{https://github.com/HuiZeng/Grid-Anchor-based-Image-Cropping-Pytorch}.
We propose K-TanH, a novel, highly accurate, hardware efficient approximation of popular activation function TanH for Deep Learning. K-TanH consists of parameterized low-precision integer operations, such as, shift and add/subtract (no floating point operation needed) where parameters are stored in very small look-up tables that can fit in CPU registers. K-TanH can work on various numerical formats, such as, Float32 and BFloat16. High quality approximations to other activation functions, e.g., Sigmoid, Swish and GELU, can be derived from K-TanH. Our AVX512 implementation of K-TanH demonstrates $>5\times$ speed up over Intel SVML, and it is consistently superior in efficiency over other approximations that use floating point arithmetic. Finally, we achieve state-of-the-art Bleu score and convergence results for training language translation model GNMT on WMT16 data sets with approximate TanH obtained via K-TanH on BFloat16 inputs.
Deep learning has become widely used in complex AI applications. Yet, training a deep neural network (DNNs) model requires a considerable amount of calculations, long running time, and much energy. Nowadays, many-core AI accelerators (e.g., GPUs and TPUs) are designed to improve the performance of AI training. However, processors from different vendors perform dissimilarly in terms of performance and energy consumption. To investigate the differences among several popular off-the-shelf processors (i.e., Intel CPU, NVIDIA GPU, AMD GPU, and Google TPU) in training DNNs, we carry out a comprehensive empirical study on the performance and energy efficiency of these processors by benchmarking a representative set of deep learning workloads, including computation-intensive operations, classical convolutional neural networks (CNNs), recurrent neural networks (LSTM), Deep Speech 2, and Transformer. Different from the existing end-to-end benchmarks which only present the training time, We try to investigate the impact of hardware, vendor's software library, and deep learning framework on the performance and energy consumption of AI training. Our evaluation methods and results not only provide an informative guide for end-users to select proper AI accelerators, but also expose some opportunities for the hardware vendors to improve their software library.
With the rapid development of in-depth learning, neural network and deep learning algorithms have been widely used in various fields, e.g., image, video and voice processing. However, the neural network model is getting larger and larger, which is expressed in the calculation of model parameters. Although a wealth of existing efforts on GPU platforms currently used by researchers for improving computing performance, dedicated hardware solutions are essential and emerging to provide advantages over pure software solutions. In this paper, we systematically investigate the neural network accelerator based on FPGA. Specifically, we respectively review the accelerators designed for specific problems, specific algorithms, algorithm features, and general templates. We also compared the design and implementation of the accelerator based on FPGA under different devices and network models and compared it with the versions of CPU and GPU. Finally, we present to discuss the advantages and disadvantages of accelerators on FPGA platforms and to further explore the opportunities for future research.
Convolutional neural networks (CNNs) have been widely deployed in the fields of computer vision and pattern recognition because of their high accuracy. However, large convolution operations are computing-intensive that often requires a powerful computing platform such as Graphics Processing Unit (GPU). This makes it difficult to apply CNNs to portable devices. The state-of-the-art CNNs, such as MobileNetV2 and Xception, adopt depthwise separable convolution to replace the standard convolution for embedded platforms. That significantly reduces operations and parameters with only limited loss in accuracy. This highly structured model is very suitable for Field-Programmable Gate Array (FPGA) implementation. In this paper, a scalable high performance depthwise separable convolution optimized CNN accelerator is proposed. The accelerator can be fit into an FPGA of different sizes, provided the balancing between hardware resources and processing speed. As an example, MobileNetV2 is implemented on Arria 10 SoC FPGA, and the results show this accelerator can classify each picture from ImageNet in 3.75ms, which is about 266.6 frames per second. This achieves 20x speedup if compared to CPU.
Tensor decomposition is one of the well-known approaches to reduce the latency time and number of parameters of a pre-trained model. However, in this paper, we propose an approach to use tensor decomposition to reduce training time of training a model from scratch. In our approach, we train the model from scratch (i.e., randomly initialized weights) with its original architecture for a small number of epochs, then the model is decomposed, and then continue training the decomposed model till the end. There is an optional step in our approach to convert the decomposed architecture back to the original architecture. We present results of using this approach on both CIFAR10 and Imagenet datasets, and show that there can be upto 2x speed up in training time with accuracy drop of upto 1.5% only, and in other cases no accuracy drop. This training acceleration approach is independent of hardware and is expected to have similar speed ups on both CPU and GPU platforms.
Nowadays most research in visual recognition using Convolutional Neural Networks (CNNs) follows the "deeper model with deeper confidence" belief to gain a higher recognition accuracy. At the same time, deeper model brings heavier computation. On the other hand, for a large chunk of recognition challenges, a system can classify images correctly using simple models or so-called shallow networks. Moreover, the implementation of CNNs faces with the size, weight, and energy constraints on the embedded devices. In this paper, we implement the adaptive switching between shallow and deep networks to reach the highest throughput on a resource-constrained MPSoC with CPU and FPGA. To this end, we develop and present a novel architecture for the CNNs where a gate makes the decision whether using the deeper model is beneficial or not. Due to resource limitation on FPGA, the idea of partial reconfiguration has been used to accommodate deep CNNs on the FPGA resources. We report experimental results on CIFAR-10, CIFAR-100, and SVHN datasets to validate our approach. Using confidence metric as the decision making factor, only 69.8%, 71.8%, and 43.8% of the computation in the deepest network is done for CIFAR-10, CIFAR-100, and SVHN while it can maintain the desired accuracy with the throughput of around 400 images per second for SVHN dataset.
The segmentation of digital images is one of the essential steps in image processing or a computer vision system. It helps in separating the pixels into different regions according to their intensity level. A large number of segmentation techniques have been proposed, and a few of them use complex computational operations. Among all, the most straightforward procedure that can be easily implemented is thresholding. In this paper, we present a unique heuristic approach for image segmentation that automatically determines multilevel thresholds by sampling the histogram of a digital image. Our approach emphasis on selecting a valley as optimal threshold values. We demonstrated that our approach outperforms the popular Otsu's method in terms of CPU computational time. We demonstrated that our approach outperforms the popular Otsu's method in terms of CPU computational time. We observed a maximum speed-up of 35.58x and a minimum speed-up of 10.21x on popular image processing benchmarks. To demonstrate the correctness of our approach in determining threshold values, we compute PSNR, SSIM, and FSIM values to compare with the values obtained by Otsu's method. This evaluation shows that our approach is comparable and better in many cases as compared to well known Otsu's method.
In this paper, we propose a novel dense surfel mapping system that scales well in different environments with only CPU computation. Using a sparse SLAM system to estimate camera poses, the proposed mapping system can fuse intensity images and depth images into a globally consistent model. The system is carefully designed so that it can build from room-scale environments to urban-scale environments using depth images from RGB-D cameras, stereo cameras or even a monocular camera. First, superpixels extracted from both intensity and depth images are used to model surfels in the system. superpixel-based surfels make our method both run-time efficient and memory efficient. Second, surfels are further organized according to the pose graph of the SLAM system to achieve $O(1)$ fusion time regardless of the scale of reconstructed models. Third, a fast map deformation using the optimized pose graph enables the map to achieve global consistency in real-time. The proposed surfel mapping system is compared with other state-of-the-art methods on synthetic datasets. The performances of urban-scale and room-scale reconstruction are demonstrated using the KITTI dataset and autonomous aggressive flights, respectively. The code is available for the benefit of the community.
Applications of Fully Convolutional Networks (FCN) in iris segmentation have shown promising advances. For mobile and embedded systems, a significant challenge is that the proposed FCN architectures are extremely computationally demanding. In this article, we propose a resource-efficient, end-to-end iris recognition flow, which consists of FCN-based segmentation, contour fitting, followed by Daugman normalization and encoding. To attain accurate and efficient FCN models, we propose a three-step SW/HW co-design methodology consisting of FCN architectural exploration, precision quantization, and hardware acceleration. In our exploration, we propose multiple FCN models, and in comparison to previous works, our best-performing model requires 50X less FLOPs per inference while achieving a new state-of-the-art segmentation accuracy. Next, we select the most efficient set of models and further reduce their computational complexity through weights and activations quantization using 8-bit dynamic fixed-point (DFP) format. Each model is then incorporated into an end-to-end flow for true recognition performance evaluation. A few of our end-to-end pipelines outperform the previous state-of-the-art on two datasets evaluated. Finally, we propose a novel DFP accelerator and fully demonstrate the SW/HW co-design realization of our flow on an embedded FPGA platform. In comparison with the embedded CPU, our hardware acceleration achieves up to 8.3X speedup for the overall pipeline while using less than 15% of the available FPGA resources. We also provide comparisons between the FPGA system and an embedded GPU showing different benefits and drawbacks for the two platforms.
In this paper, we present a generic and robust multimodal synthesis system that produces highly natural speech and facial expression simultaneously. The key component of this system is the Duration Informed Attention Network (DurIAN), an autoregressive model in which the alignments between the input text and the output acoustic features are inferred from a duration model. This is different from the end-to-end attention mechanism used, and accounts for various unavoidable artifacts, in existing end-to-end speech synthesis systems such as Tacotron. Furthermore, DurIAN can be used to generate high quality facial expression which can be synchronized with generated speech with/without parallel speech and face data. To improve the efficiency of speech generation, we also propose a multi-band parallel generation strategy on top of the WaveRNN model. The proposed Multi-band WaveRNN effectively reduces the total computational complexity from 9.8 to 5.5 GFLOPS, and is able to generate audio that is 6 times faster than real time on a single CPU core. We show that DurIAN could generate highly natural speech that is on par with current state of the art end-to-end systems, while at the same time avoid word skipping/repeating errors in those systems. Finally, a simple yet effective approach for fine-grained control of expressiveness of speech and facial expression is introduced.
This paper presents a fast projection mapping method for moving image content projected onto a markerless planar surface using a low-latency Digital Micromirror Device (DMD) projector. By adopting a closed-loop alignment approach, in which not only the surface texture but also the projected image is tracked by a camera, the proposed method is free from a calibration or position adjustment between the camera and projector. We designed fiducial patterns to be inserted into a fast flapping sequence of binary frames of the DMD projector, which allows the simultaneous tracking of the surface texture and a fiducial geometry separate from a single image captured by the camera. The proposed method implemented on a CPU runs at 400 fps and enables arbitrary video contents to be "stuck" onto a variety of textured surfaces.
We propose a practical scheme to train a single multilingual sequence labeling model that yields state of the art results and is small and fast enough to run on a single CPU. Starting from a public multilingual BERT checkpoint, our final model is 6x smaller and 27x faster, and has higher accuracy than a state-of-the-art multilingual baseline. We show that our model especially outperforms on low-resource languages, and works on codemixed input text without being explicitly trained on codemixed examples. We showcase the effectiveness of our method by reporting on part-of-speech tagging and morphological prediction on 70 treebanks and 48 languages.