The condition assessment of road surfaces is essential to ensure their serviceability while still providing maximum road traffic safety. This paper presents a robust stereo vision system embedded in an unmanned aerial vehicle (UAV). The perspective view of the target image is first transformed into the reference view, and this not only improves the disparity accuracy, but also reduces the algorithm's computational complexity. The cost volumes generated from stereo matching are then filtered using a bilateral filter. The latter has been proved to be a feasible solution for the functional minimisation problem in a fully connected Markov random field model. Finally, the disparity maps are transformed by minimising an energy function with respect to the roll angle and disparity projection model. This makes the damaged road areas more distinguishable from the road surface. The proposed system is implemented on an NVIDIA Jetson TX2 GPU with CUDA for real-time purposes. It is demonstrated through experiments that the damaged road areas can be easily distinguished from the transformed disparity maps.
The performance of video saliency estimation techniques has achieved significant advances along with the rapid development of Convolutional Neural Networks (CNNs). However, devices like cameras and drones may have limited computational capability and storage space so that the direct deployment of complex deep saliency models becomes infeasible. To address this problem, this paper proposes a dynamic saliency estimation approach for aerial videos via spatiotemporal knowledge distillation. In this approach, five components are involved, including two teachers, two students and the desired spatiotemporal model. The knowledge of spatial and temporal saliency is first separately transferred from the two complex and redundant teachers to their simple and compact students, and the input scenes are also degraded from high-resolution to low-resolution to remove the probable data redundancy so as to greatly speed up the feature extraction process. After that, the desired spatiotemporal model is further trained by distilling and encoding the spatial and temporal saliency knowledge of two students into a unified network. In this manner, the inter-model redundancy can be further removed for the effective estimation of dynamic saliency on aerial videos. Experimental results show that the proposed approach outperforms ten state-of-the-art models in estimating visual saliency on aerial videos, while its speed reaches up to 28,738 FPS on the GPU platform.
Road scene understanding is a critical component in an autonomous driving system. Although the deep learning-based road scene segmentation can achieve very high accuracy, its complexity is also very high for developing real-time applications. It is challenging to design a neural net with high accuracy and low computational complexity. To address this issue, we investigate the advantages and disadvantages of several popular CNN architectures in terms of speed, storage and segmentation accuracy. We start from the Fully Convolutional Network (FCN) with VGG, and then we study ResNet and DenseNet. Through detailed experiments, we pick up the favorable components from the existing architectures and at the end, we construct a light-weight network architecture based on the DenseNet. Our proposed network, called DSNet, demonstrates a real-time testing (inferencing) ability (on the popular GPU platform) and it maintains an accuracy comparable with most previous systems. We test our system on several datasets including the challenging Cityscapes dataset (resolution of 1024x512) with an mIoU of about 69.1 % and runtime of 0.0147 second per image on a single GTX 1080Ti. We also design a more accurate model but at the price of a slower speed, which has an mIoU of about 72.6 % on the CamVid dataset.
While embedded FPGAs are attractive platforms for DNN acceleration on edge-devices due to their low latency and high energy efficiency, the scarcity of resources of edge-scale FPGA devices also makes it challenging for DNN deployment. In this paper, we propose a simultaneous FPGA/DNN co-design methodology with both bottom-up and top-down approaches: a bottom-up hardware-oriented DNN model search for high accuracy, and a top-down FPGA accelerator design considering DNN-specific characteristics. We also build an automatic co-design flow, including an Auto-DNN engine to perform hardware-oriented DNN model search, as well as an Auto-HLS engine to generate synthesizable C code of the FPGA accelerator for explored DNNs. We demonstrate our co-design approach on an object detection task using PYNQ-Z1 FPGA. Results show that our proposed DNN model and accelerator outperform the state-of-the-art FPGA designs in all aspects including Intersection-over-Union (IoU) (6.2% higher), frames per second (FPS) (2.48X higher), power consumption (40% lower), and energy efficiency (2.5X higher). Compared to GPU-based solutions, our designs deliver similar accuracy but consume far less energy.
Recommender systems that can learn from cross-session data to dynamically predict the next item a user will choose are crucial for online platforms. However, existing approaches often use out-of-the-box sequence models which are limited by speed and memory consumption, are often infeasible for production environments, and usually do not incorporate cross-session information, which is crucial for effective recommendations. Here we propose Hierarchical Temporal Convolutional Networks (HierTCN), a hierarchical deep learning architecture that makes dynamic recommendations based on users' sequential multi-session interactions with items. HierTCN is designed for web-scale systems with billions of items and hundreds of millions of users. It consists of two levels of models: The high-level model uses Recurrent Neural Networks (RNN) to aggregate users' evolving long-term interests across different sessions, while the low-level model is implemented with Temporal Convolutional Networks (TCN), utilizing both the long-term interests and the short-term interactions within sessions to predict the next interaction. We conduct extensive experiments on a public XING dataset and a large-scale Pinterest dataset that contains 6 million users with 1.6 billion interactions. We show that HierTCN is 2.5x faster than RNN-based models and uses 90% less data memory compared to TCN-based models. We further develop an effective data caching scheme and a queue-based mini-batch generator, enabling our model to be trained within 24 hours on a single GPU. Our model consistently outperforms state-of-the-art dynamic recommendation methods, with up to 18% improvement in recall and 10% in mean reciprocal rank.
The greatest challenge facing visual object tracking is the simultaneous requirements on robustness and discrimination power. In this paper, we propose a SiamFC-based tracker, named SPM-Tracker, to tackle this challenge. The basic idea is to address the two requirements in two separate matching stages. Robustness is strengthened in the coarse matching (CM) stage through generalized training while discrimination power is enhanced in the fine matching (FM) stage through a distance learning network. The two stages are connected in series as the input proposals of the FM stage are generated by the CM stage. They are also connected in parallel as the matching scores and box location refinements are fused to generate the final results. This innovative series-parallel structure takes advantage of both stages and results in superior performance. The proposed SPM-Tracker, running at 120fps on GPU, achieves an AUC of 0.687 on OTB-100 and an EAO of 0.434 on VOT-16, exceeding other real-time trackers by a notable margin.
Large convolutional neural network models have recently demonstrated impressive performance on video attention prediction. Conventionally, these models are with intensive computation and large memory. To address these issues, we design an extremely light-weight network with ultrafast speed, named UVA-Net. The network is constructed based on depth-wise convolutions and takes low-resolution images as input. However, this straight-forward acceleration method will decrease performance dramatically. To this end, we propose a coupled knowledge distillation strategy to augment and train the network effectively. With this strategy, the model can further automatically discover and emphasize implicit useful cues contained in the data. Both spatial and temporal knowledge learned by the high-resolution complex teacher networks also can be distilled and transferred into the proposed low-resolution light-weight spatiotemporal network. Experimental results show that the performance of our model is comparable to 11 state-of-the-art models in video attention prediction, while it costs only 0.68 MB memory footprint, runs about 10,106 FPS on GPU and 404 FPS on CPU, which is 206 times faster than previous models.
Stack Long Short-Term Memory (StackLSTM) is useful for various applications such as parsing and string-to-tree neural machine translation, but it is also known to be notoriously difficult to parallelize for GPU training due to the fact that the computations are dependent on discrete operations. In this paper, we tackle this problem by utilizing state access patterns of StackLSTM to homogenize computations with regard to different discrete operations. Our parsing experiments show that the method scales up almost linearly with increasing batch size, and our parallelized PyTorch implementation trains significantly faster compared to the Dynet C++ implementation.
Over the past few years machine learning has seen a renewed explosion of interest, following a number of studies showing the effectiveness of neural networks in a range of tasks which had previously been considered incredibly hard. Neural networks' effectiveness in the fields of image recognition and natural language processing stems primarily from the vast amounts of data available to companies and researchers, coupled with the huge amounts of compute power available in modern accelerators such as GPUs, FPGAs and ASICs. There are a number of approaches available to developers for utilizing GPGPU technologies such as SYCL, OpenCL and CUDA, however many applications require the same low level mathematical routines. Libraries dedicated to accelerating these common routines allow developers to easily make full use of the available hardware without requiring low level knowledge of the hardware themselves, however such libraries are often provided by hardware manufacturers for specific hardware such as cuDNN for Nvidia hardware or MIOpen for AMD hardware. SYCL-DNN is a new open-source library dedicated to providing accelerated routines for neural network operations which are hardware and vendor agnostic. Built on top of the SYCL open standard and written entirely in standard C++, SYCL-DNN allows a user to easily accelerate neural network code for a wide range of hardware using a modern C++ interface. The library is tested on AMD's OpenCL for GPU, Intel's OpenCL for CPU and GPU, ARM's OpenCL for Mali GPUs as well as ComputeAorta's OpenCL for R-Car CV engine and host CPU. In this talk we will present performance figures for SYCL-DNN on this range of hardware, and discuss how high performance was achieved on such a varied set of accelerators with such different hardware features.
Automatic methods for Neural Architecture Search (NAS) have been shown to produce state-of-the-art network models. Yet, their main drawback is the computational complexity of the search process. As some primal methods optimized over a discrete search space, thousands of days of GPU were required for convergence. A recent approach is based on constructing a differentiable search space that enables gradient-based optimization, which reduces the search time to a few days. While successful, it still includes some noncontinuous steps, e.g., the pruning of many weak connections at once. In this paper, we propose a differentiable search space that allows the annealing of architecture weights, while gradually pruning inferior operations. In this way, the search converges to a single output network in a continuous manner. Experiments on several vision datasets demonstrate the effectiveness of our method with respect to the search cost and accuracy of the achieved model. Specifically, with $0.2$ GPU search days we achieve an error rate of $1.68\%$ on CIFAR-10.
Can we automatically design a Convolutional Network (ConvNet) with the highest image classification accuracy under the runtime constraint of a mobile device? Neural architecture search (NAS) has revolutionized the design of hardware-efficient ConvNets by automating this process. However, the NAS problem remains challenging due to the combinatorially large design space, causing a significant searching time (at least 200 GPU-hours). To alleviate this complexity, we propose Single-Path NAS, a novel differentiable NAS method for designing hardware-efficient ConvNets in less than 4 hours. Our contributions are as follows: 1. Single-path search space: Compared to previous differentiable NAS methods, Single-Path NAS uses one single-path over-parameterized ConvNet to encode all architectural decisions with shared convolutional kernel parameters, hence drastically decreasing the number of trainable parameters and the search cost down to few epochs. 2. Hardware-efficient ImageNet classification: Single-Path NAS achieves 74.96% top-1 accuracy on ImageNet with 79ms latency on a Pixel 1 phone, which is state-of-the-art accuracy compared to NAS methods with similar constraints (<80ms). 3. NAS efficiency: Single-Path NAS search cost is only 8 epochs (30 TPU-hours), which is up to 5,000x faster compared to prior work. 4. Reproducibility: Unlike all recent mobile-efficient NAS methods which only release pretrained models, we open-source our entire codebase at: https://github.com/dstamoulis/single-path-nas.
In many real-world planning problems with factored, mixed discrete and continuous state and action spaces such as Reservoir Control, Heating Ventilation, and Air Conditioning, and Navigation domains, it is difficult to obtain a model of the complex nonlinear dynamics that govern state evolution. However, the ubiquity of modern sensors allows us to collect large quantities of data from each of these complex systems and build accurate, nonlinear deep neural network models of their state transitions. But there remains one major problem for the task of control -- how can we plan with deep network learned transition models without resorting to Monte Carlo Tree Search and other black-box transition model techniques that ignore model structure and do not easily extend to mixed discrete and continuous domains? In this paper, we introduce two types of nonlinear planning methods that can leverage deep neural network learned transition models: Hybrid Deep MILP Planner (HD-MILP-Plan) and Tensorflow Planner (TF-Plan). In HD-MILP-Plan, we make the critical observation that the Rectified Linear Unit transfer function for deep networks not only allows faster convergence of model learning, but also permits a direct compilation of the deep network transition model to a Mixed-Integer Linear Program encoding. Further, we identify deep network specific optimizations for HD-MILP-Plan that improve performance over a base encoding and show that we can plan optimally with respect to the learned deep networks. In TF-Plan, we take advantage of the efficiency of auto-differentiation tools and GPU-based computation where we encode a subclass of purely continuous planning problems as Recurrent Neural Networks and directly optimize the actions through backpropagation. We compare both planners and show that TF-Plan is able to approximate the optimal plans found by HD-MILP-Plan in less computation time...
We present a simple, fully-convolutional model for real-time instance segmentation that achieves 29.8 mAP on MS COCO at 33.5 fps evaluated on a single Titan Xp, which is significantly faster than any previous competitive approach. Moreover, we obtain this result after training on only one GPU. We accomplish this by breaking instance segmentation into two parallel subtasks: (1) generating a set of prototype masks and (2) predicting per-instance mask coefficients. Then we produce instance masks by linearly combining the prototypes with the mask coefficients. We find that because this process doesn't depend on repooling, this approach produces very high-quality masks and exhibits temporal stability for free. Furthermore, we analyze the emergent behavior of our prototypes and show they learn to localize instances on their own in a translation variant manner, despite being fully-convolutional. Finally, we also propose Fast NMS, a drop-in 12 ms faster replacement for standard NMS that only has a marginal performance penalty.
Recently, convolutional neural networks with 3D kernels (3D CNNs) have been very popular in computer vision community as a result of their superior ability of extracting spatio-temporal features within video frames compared to 2D CNNs. Although there has been great advances recently to build resource efficient 2D CNN architectures considering memory and power budget, there is hardly any similar resource efficient architectures for 3D CNNs. In this paper, we have converted various well-known resource efficient 2D CNNs to 3D CNNs and evaluated their performance on three major benchmarks in terms of classification accuracy for different complexity levels. We have experimented on (1) Kinetics-600 dataset to inspect their capacity to learn, (2) Jester dataset to inspect their ability to capture motion patterns, and (3) UCF-101 to inspect the applicability of transfer learning. We have evaluated the run-time performance of each model on a single Titan XP GPU and a Jetson TX2 embedded system. The results of this study show that these models can be utilized for different types of real-world applications since they provide real-time performance with considerable accuracies and memory usage. Our analysis on different complexity levels shows that the resource efficient 3D CNNs should not be designed too shallow or narrow in order to save complexity. The codes and pretrained models used in this work are publicly available.
In semantic video segmentation the goal is to acquire consistent dense semantic labelling across image frames. To this end, recent approaches have been reliant on manually arranged operations applied on top of static semantic segmentation networks - with the most prominent building block being the optical flow able to provide information about scene dynamics. Related to that is the line of research concerned with speeding up static networks by approximating expensive parts of them with cheaper alternatives, while propagating information from previous frames. In this work we attempt to come up with generalisation of those methods, and instead of manually designing contextual blocks that connect per-frame outputs, we propose a neural architecture search solution, where the choice of operations together with their sequential arrangement are being predicted by a separate neural network. We showcase that such generalisation leads to stable and accurate results across common benchmarks, such as CityScapes and CamVid datasets. Importantly, the proposed methodology takes only 2 GPU-days, finds high-performing cells and does not rely on the expensive optical flow computation.
Training convolutional neural networks (CNNs) requires intense compute throughput and high memory bandwidth. Especially, convolution layers account for the majority of the execution time of CNN training, and GPUs are commonly used to accelerate these layer workloads. GPU design optimization for efficient CNN training acceleration requires the accurate modeling of how their performance improves when computing and memory resources are increased. We present DeLTA, the first analytical model that accurately estimates the traffic at each GPU memory hierarchy level, while accounting for the complex reuse patterns of a parallel convolution algorithm. We demonstrate that our model is both accurate and robust for different CNNs and GPU architectures. We then show how this model can be used to carefully balance the scaling of different GPU resources for efficient CNN performance improvement.
We present Habitat, a platform for research in embodied artificial intelligence (AI). Habitat enables training embodied agents (virtual robots) in highly efficient photorealistic 3D simulation. Specifically, Habitat consists of: (i) Habitat-Sim: a flexible, high-performance 3D simulator with configurable agents, sensors, and generic 3D dataset handling. Habitat-Sim is fast -- when rendering a scene from Matterport3D, it achieves several thousand frames per second (fps) running single-threaded, and can reach over 10,000 fps multi-process on a single GPU. (ii) Habitat-API: a modular high-level library for end-to-end development of embodied AI algorithms -- defining tasks (e.g., navigation, instruction following, question answering), configuring, training, and benchmarking embodied agents. These large-scale engineering contributions enable us to answer scientific questions requiring experiments that were till now impracticable or 'merely' impractical. Specifically, in the context of point-goal navigation: (1) we revisit the comparison between learning and SLAM approaches from two recent works and find evidence for the opposite conclusion -- that learning outperforms SLAM if scaled to an order of magnitude more experience than previous investigations, and (2) we conduct the first cross-dataset generalization experiments {train, test} x {Matterport3D, Gibson} for multiple sensors {blind, RGB, RGBD, D} and find that only agents with depth (D) sensors generalize across datasets. We hope that our open-source platform and these findings will advance research in embodied AI.
In this paper, we present a short description of the method proposed to ANHIR challenge organized jointly with the IEEE ISBI 2019 conference. We propose a method consisting of preprocessing, initial alignment, nonrigid registration algorithms and a method to automatically choose the best result. The method turned out to be robust (99.792% robustness) and accurate (0.38% average median rTRE). The main drawback of the proposed method is relatively high computation time. However, this aspect can be easily improved by cleaning the code and proposing a GPU implementation.
The use of large-scale multifaceted data is common in a wide variety of scientific applications. In many cases, this multifaceted data takes the form of a field-based (Eulerian) and point/trajectory-based (Lagrangian) representation as each has a unique set of advantages in characterizing a system of study. Furthermore, studying the increasing scale and complexity of these multifaceted datasets is limited by perceptual ability and available computational resources, necessitating sophisticated data reduction and feature extraction techniques. In this work, we present a new 4D feature segmentation/extraction scheme that can operate on both the field and point/trajectory data types simultaneously. The resulting features are time-varying data subsets that have both a field and point-based component, and were extracted based on underlying patterns from both data types. This enables researchers to better explore both the spatial and temporal interplay between the two data representations and study underlying phenomena from new perspectives. We parallelize our approach using GPU acceleration and apply it to real world multifaceted datasets to illustrate the types of features that can be extracted and explored.
Detailed whole brain segmentation is an essential quantitative technique, which provides a non-invasive way of measuring brain regions from a structural magnetic resonance imaging (MRI). Recently, deep convolution neural network (CNN) has been applied to whole brain segmentation. However, restricted by current GPU memory, 2D based methods, downsampling based 3D CNN methods, and patch-based high-resolution 3D CNN methods have been the de facto standard solutions. 3D patch-based high resolution methods typically yield superior performance among CNN approaches on detailed whole brain segmentation (>100 labels), however, whose performance are still commonly inferior compared with multi-atlas segmentation methods (MAS) due to the following challenges: (1) a single network is typically used to learn both spatial and contextual information for the patches, (2) limited manually traced whole brain volumes are available (typically less than 50) for training a network. In this work, we propose the spatially localized atlas network tiles (SLANT) method to distribute multiple independent 3D fully convolutional networks (FCN) for high-resolution whole brain segmentation. To address the first challenge, multiple spatially distributed networks were used in the SLANT method, in which each network learned contextual information for a fixed spatial location. To address the second challenge, auxiliary labels on 5111 initially unlabeled scans were created by multi-atlas segmentation for training. Since the method integrated multiple traditional medical image processing methods with deep learning, we developed a containerized pipeline to deploy the end-to-end solution. From the results, the proposed method achieved superior performance compared with multi-atlas segmentation methods, while reducing the computational time from >30 hours to 15 minutes (https://github.com/MASILab/SLANTbrainSeg).