Search

Deep learning models have been demonstrated vulnerable to adversarial attacks even with imperceptible perturbations. As such, the reliability of existing deep neural networks-based autonomous driving systems can suffer. However, deep 3D models have applications in various Cyber-Physical Systems (CPSs) with safety-critical requirements, particularly autonomous driving systems. In this paper, the robustness of deep 3D object detection models under adversarial point cloud perturbations has been investigated. A novel method is developed to generate 3D adversarial examples from point cloud perturbations, which are common due to the intrinsic characteristics of the data captured by 3D sensors, e.g., LiDAR. The generation of adversarial samples is supervised by a dual loss, which constitutes an adversarial loss and a perturbation loss. The adversarial loss produces a point cloud with the property of aggressiveness, while the perturbation loss enforces the produced point cloud subject to visual imperception. We demonstrate that the method can successfully attack 3D object detection models in most cases, and expose their vulnerability to physical-world attacks in the form of point cloud perturbations. We perform a thorough evaluation of popular deep 3D object detectors in an adversarial setting on the KITTI vision benchmark. Experimental results show that current deep 3D object detection models are susceptible to adversarial attacks in the context of autonomous driving, and their performances are degraded by a large margin in the presence of adversarial point clouds generated by the proposed method.

We consider the 3D object recognition problem from the perspective of the lack of labelled data. In this paper, we propose a novel progressive conditional generative adversarial network (PC-GAN) for 3D object recognition by conditioning the input with progressive learning strategies. PC-GAN is a powerful adversarial model whose generator automatically produces realistic 3D objects with annotations, and the discriminator distinguishes them from the training distribution and recognizes their categories. We train the discriminative classifier simultaneously with the generator to predict the class label by embedding a SoftMax classifier. Progressive learning uses input samples from lower to higher resolutions to increase the generator performance gradually and produce informative objects for a certain class of objects. The key idea of adopting progressing learning is to mitigate overshoots issues of the discriminator and increase variations in the generated objects by learning progressively. This strategy helps the generator to produce more realistic synthetic objects and improve the active classification performance of the discriminator. Our proposed PC-GAN is trained for object classification in a supervised manner and the performance is evaluated on two public datasets. Experimental results demonstrate that our adversarial PC-GAN outperforms the existing volumetric discriminative classifiers in term of classification accuracy.

Missing data is a common problem faced with real-world datasets. Imputation is a widely used technique to estimate the missing data. State-of-the-art imputation approaches model the distribution of observed data to approximate the missing values. Such an approach usually models a single distribution for the entire dataset, which overlooks the class-specific characteristics of the data. Class-specific characteristics are especially useful when there is a class imbalance. We propose a new method for imputing missing data based on its class-specific characteristics by adapting the popular Conditional Generative Adversarial Networks (CGAN). Our Conditional Generative Adversarial Imputation Network (CGAIN) imputes the missing data using class-specific distributions, which can produce the best estimates for the missing values. We tested our approach on baseline datasets and achieved superior performance compared with the state-of-the-art and popular imputation approaches.

Weakly supervised semantic segmentation has been attracting increasing attention as it can alleviate the need for expensive pixel-level annotations through the use of image-level labels. Relevant methods mainly rely on the implicit object localization ability of convolutional neural networks (CNNs). However, generated object attention maps remain mostly small and incomplete. In this paper, we propose an Atrous Convolutional Feature Network (ACFN) to generate dense object attention maps. This is achieved by enhancing the context representation of image classification CNNs. More specifically, cascaded atrous convolutions are used in the middle layers to retain sufficient spatial details, and pyramidal atrous convolutions are used in the last convolutional layers to provide multi-scale context information for the extraction of object attention maps. Moreover, we propose an attentive fusion strategy to adaptively fuse the multi-scale features. Our method shows improvements over existing methods on both the PASCAL VOC 2012 and MS COCO datasets, achieving state-of-the-art performance.

Oceanographers rely on advanced digital imaging systems to assess the health of marine ecosystems. The majority of the imagery collected by these systems do not get annotated due to lack of resources. Consequently, the expert labeled data is not enough to train dedicated deep networks. Meanwhile, in the deep learning community, much focus is on how to use pre-trained deep networks to classify out-of-domain images and transfer learning. In this paper, we leverage these advances to evaluate how well features extracted from deep neural networks transfer to underwater image classification. We propose new image features (called ResFeats) extracted from the different convolutional layers of a deep residual network pre-trained on ImageNet. We further combine the ResFeats extracted from different layers to obtain compact and powerful deep features. Moreover, we show that ResFeats consistently perform better than their CNN counterparts. Experimental results are provided to show the effectiveness of ResFeats with state-of-the-art classification accuracies on MLC, Benthoz15, EILAT and RSMAS datasets.

A common approach to tackle 3D object recognition tasks is to project 3D data to multiple 2D images. Projection only captures the outline of the object, and discards the internal information that may be crucial for the recognition. In this paper, we stay in 3D and concentrate on tapping the potential of 3D representations. We present NormalNet, a voxel-based convolutional neural network (CNN) designed for 3D object recognition. The network uses normal vectors of the object surfaces as input, which demonstrate stronger discrimination capability than binary voxels. We propose a reflection-convolution-concatenation (RCC) module to realize the conv layers, which extracts distinguishable features for 3D vision tasks while reducing the number of parameters significantly. We further improve the performance of NormalNet by combining two networks, which take normal vectors and voxels as input respectively. We carry out a series of experiments that validate the design of the network and achieve competitive performance in 3D object classification and retrieval tasks.

This paper presents a learned feature based framework for both outdoor and indoor scene labeling. This framework is combined with a discriminative feature learning process to produce the posteriors of every pixel and a novel strategy to improve the global label consistency of a scene. First, we use Convolutional Neural Networks (ConvNets) to learn the most relevant features of a scene at the multi-scale superpixel level. The effect of both trained and general ConvNets features for our scene labeling framework are investigated. Then, based on the predicted posteriors from the learned features, we propose an algorithm called Region Consistency Activation (RCA) to iteratively improve the global label consistency at different levels of the Ultrametric Contour Map (UCM). In addition, we propose a strategy to make the hyper-parameters of RCA adaptive to the test images, which results in a better generalization ability compared with the hyper-parameters tuning based RCA. Our scene labeling framework were rigorously tested on three popular scene labeling datasets: Stanford Background, SIFT Flow and NYU-Depth V2. Experiments show that our proposed method consistently produces better accuracy and visual consistency compared with the state-of-the-art methods for both outdoor and indoor scenes.