ICCV 2019 论文列表

Deep neural networks have lead to a breakthrough in depth estimation from single images. Recent work shows that the quality of these estimations is rapidly increasing. It is clear that neural networks can see depth in single images. However, to the best of our knowledge, no work currently exists that analyzes what these networks have learned. In this work we take four previously published networks and investigate what depth cues they exploit. We find that all networks ignore the apparent size of known obstacles in favor of their vertical position in the image. The use of the vertical position requires the camera pose to be known; however, we find that these networks only partially recognize changes in camera pitch and roll angles. Small changes in camera pitch are shown to disturb the estimated distance towards obstacles. The use of the vertical image position allows the networks to estimate depth towards arbitrary obstacles - even those not appearing in the training set - but may depend on features that are not universally present.

This paper proposes a geometric interpretation of the angles and scales which the orientation- and scale-covariant feature detectors, e.g. SIFT, provide. Two new general constraints are derived on the scales and rotations which can be used in any geometric model estimation tasks. Using these formulas, two new constraints on homography estimation are introduced. Exploiting the derived equations, a solver for estimating the homography from the minimal number of two correspondences is proposed. Also, it is shown how the normalization of the point correspondences affects the rotation and scale parameters, thus achieving numerically stable results. Due to requiring merely two feature pairs, robust estimators, e.g. RANSAC, do significantly fewer iterations than by using the four-point algorithm. When using covariant features, e.g. SIFT, this additional information is given at no cost. The method is tested in a synthetic environment and on publicly available real-world datasets.

We propose a novel generative adversarial network (GAN) for the task of unsupervised learning of 3D representations from natural images. Most generative models rely on 2D kernels to generate images and make few assumptions about the 3D world. These models therefore tend to create blurry images or artefacts in tasks that require a strong 3D understanding, such as novel-view synthesis. HoloGAN instead learns a 3D representation of the world, and to render this representation in a realistic manner. Unlike other GANs, HoloGAN provides explicit control over the pose of generated objects through rigid-body transformations of the learnt 3D features. Our experiments show that using explicit 3D features enables HoloGAN to disentangle 3D pose and identity, which is further decomposed into shape and appearance, while still being able to generate images with similar or higher visual quality than other generative models. HoloGAN can be trained end-to-end from unlabelled 2D images only. Particularly, we do not require pose labels, 3D shapes, or multiple views of the same objects. This shows that HoloGAN is the first generative model that learns 3D representations from natural images in an entirely unsupervised manner.

We propose a new 3D holistic++ scene understanding problem, which jointly tackles two tasks from a single-view image: (i) holistic scene parsing and reconstruction---3D estimations of object bounding boxes, camera pose, and room layout, and (ii) 3D human pose estimation. The intuition behind is to leverage the coupled nature of these two tasks to improve the granularity and performance of scene understanding. We propose to exploit two critical and essential connections between these two tasks: (i) human-object interaction (HOI) to model the fine-grained relations between agents and objects in the scene, and (ii) physical commonsense to model the physical plausibility of the reconstructed scene. The optimal configuration of the 3D scene, represented by a parse graph, is inferred using Markov chain Monte Carlo (MCMC), which efficiently traverses through the non-differentiable joint solution space. Experimental results demonstrate that the proposed algorithm significantly improves the performance of the two tasks on three datasets, showing an improved generalization ability.

Adversarial perturbations of clean images are usually imperceptible for human eyes, but can confidently fool deep neural networks (DNNs) to make incorrect predictions. Such vulnerability of DNNs raises serious security concerns about their practicability in security-sensitive applications. To defend against such adversarial perturbations, recently developed PixelDefend purifies a perturbed image based on PixelCNN in a raster scan order (row/column by row/column). However, such scan mode insufficiently exploits the correlations between pixels, which further limits its robustness performance. Therefore, we propose a more advanced Hilbert curve scan order to model the pixel dependencies in this paper. Hilbert curve could well preserve local consistency when mapping from 2-D image to 1-D vector, thus the local features in neighboring pixels can be more effectively modeled. Moreover, the defensive power can be further improved via ensembles of Hilbert curve with different orientations. Experimental results demonstrate the superiority of our method over the state-of-the-art defenses against various adversarial attacks.

It is always well believed that parsing an image into constituent visual patterns would be helpful for understanding and representing an image. Nevertheless, there has not been evidence in support of the idea on describing an image with a natural-language utterance. In this paper, we introduce a new design to model a hierarchy from instance level (segmentation), region level (detection) to the whole image to delve into a thorough image understanding for captioning. Specifically, we present a HIerarchy Parsing (HIP) architecture that novelly integrates hierarchical structure into image encoder. Technically, an image decomposes into a set of regions and some of the regions are resolved into finer ones. Each region then regresses to an instance, i.e., foreground of the region. Such process naturally builds a hierarchal tree. A tree-structured Long Short-Term Memory (Tree-LSTM) network is then employed to interpret the hierarchal structure and enhance all the instance-level, region-level and image-level features. Our HIP is appealing in view that it is pluggable to any neural captioning models. Extensive experiments on COCO image captioning dataset demonstrate the superiority of HIP. More remarkably, HIP plus a top-down attention-based LSTM decoder increases CIDEr-D performance from 120.1% to 127.2% on COCO Karpathy test split. When further endowing instance-level and region-level features from HIP with semantic relation learnt through Graph Convolutional Networks (GCN), CIDEr-D is boosted up to 130.6%.

Single shot detector simultaneously predicts object categories and regression offsets of the default boxes. Despite of high efficiency, this structure has some inappropriate designs: (1) The classification result of the default box is improperly assigned to that of the regressed box during inference, (2) Only regression once is not good enough for accurate object detection. To solve the first problem, a novel reg-offset-cls (ROC) module is proposed. It contains three hierarchical steps: box regression, the feature sampling location predication, and the regressed box classification with the features of offset locations. To further solve the second problem, a hierarchical shot detector (HSD) is proposed, which stacks two ROC modules and one feature enhanced module. The second ROC treats the regressed boxes and the feature sampling locations of features in the first ROC as the inputs. Meanwhile, the feature enhanced module injected between two ROCs aims to extract the local and non-local context. Experiments on the MS COCO and PASCAL VOC datasets demonstrate the superiority of proposed HSD. Without the bells or whistles, HSD outperforms all one-stage methods at real-time speed.

This paper presents a novel Hierarchical Self-Attention Network (HISAN) to generate spatial-temporal tubes for action localization in videos. The essence of HISAN is to combine the two-stream convolutional neural network (CNN) with hierarchical bidirectional self-attention mechanism, which comprises of two levels of bidirectional self-attention to efficaciously capture both of the long-term temporal dependency information and spatial context information to render more precise action localization. Also, a sequence rescoring (SR) algorithm is employed to resolve the dilemma of inconsistent detection scores incurred by occlusion or background clutter. Moreover, a new fusion scheme is invoked, which integrates not only the appearance and motion information from the two-stream network, but also the motion saliency to mitigate the effect of camera motion. Simulations reveal that the new approach achieves competitive performance as the state-of-the-art works in terms of action localization and recognition accuracy on the widespread UCF101-24 and J-HMDB datasets.

We achieve 3D semantic scene labeling by exploring semantic relation between each point and its contextual neighbors through edges. Besides an encoder-decoder branch for predicting point labels, we construct an edge branch to hierarchically integrate point features and generate edge features. To incorporate point features in the edge branch, we establish a hierarchical graph framework, where the graph is initialized from a coarse layer and gradually enriched along the point decoding process. For each edge in the final graph, we predict a label to indicate the semantic consistency of the two connected points to enhance point prediction. At different layers, edge features are also fed into the corresponding point module to integrate contextual information for message passing enhancement in local regions. The two branches interact with each other and cooperate in segmentation. Decent experimental results on several 3D semantic labeling datasets demonstrate the effectiveness of our work.

Snapshot-based visual localization is an important problem in several computer vision and robotics applications such as Simultaneous Localization And Mapping (SLAM). To achieve real-time performance in very large-scale environments with massive amounts of training and map data, techniques such as approximate nearest neighbor search (ANN) algorithms are used. While several state-of-the-art variants of quantization and indexing techniques have demonstrated to be efficient in practice, their theoretical memory cost still scales at least linearly with the training data (i.e., O(n) where n is the number of instances in the database), since each data point must be associated with at least one code vector. To address these limitations, in this paper we present a totally new hierarchical encoding approach that enables a sub-linear storage scale. The algorithm exploits the widespread sequential nature of sensor information streams in robotics and autonomous vehicle applications and achieves, both theoretically and experimentally, sub-linear scalability in storage required for a given environment size. Furthermore, the associated query time of our algorithm is also of sub-linear complexity. We benchmark the performance of the proposed algorithm on several real-world benchmark datasets and experimentally validate the theoretical sub-linearity of our approach, while also showing that our approach yields competitive absolute storage performance as well.

We present a method for hiding video content inside audio files while preserving the perceptual fidelity of the cover audio. This is a form of cross-modal steganography and is particularly challenging due to the high bitrate of video. Our scheme uses recent advances in flow-based generative models, which enable mapping audio to latent codes such that nearby codes correspond to perceptually similar signals. We show that compressed video data can be concealed in the latent codes of audio sequences while preserving the fidelity of both the hidden video and the cover audio. We can embed 128x128 video inside same-duration audio, or higher-resolution video inside longer audio sequences. Quantitative experiments show that our approach outperforms relevant baselines in steganographic capacity and fidelity.

The matching of multiple objects (e.g. shapes or images) is a fundamental problem in vision and graphics. In order to robustly handle ambiguities, noise and repetitive patterns in challenging real-world settings, it is essential to take geometric consistency between points into account. Computationally, the multi-matching problem is difficult. It can be phrased as simultaneously solving multiple (NP-hard) quadratic assignment problems (QAPs) that are coupled via cycle-consistency constraints. The main limitations of existing multi-matching methods are that they either ignore geometric consistency and thus have limited robustness, or they are restricted to small-scale problems due to their (relatively) high computational cost. We address these shortcomings by introducing a Higher-order Projected Power Iteration method, which is (i) efficient and scales to tens of thousands of points, (ii) straightforward to implement, (iii) able to incorporate geometric consistency, (iv) guarantees cycle-consistent multi-matchings, and (iv) comes with theoretical convergence guarantees. Experimentally we show that our approach is superior to existing methods.

State-of-the-art neural network architectures such as ResNet, MobileNet, and DenseNet have achieved outstanding accuracy over low MACs and small model size counterparts. However, these metrics might not be accurate for predicting the inference time. We suggest that memory traffic for accessing intermediate feature maps can be a factor dominating the inference latency, especially in such tasks as real-time object detection and semantic segmentation of high-resolution video. We propose a Harmonic Densely Connected Network to achieve high efficiency in terms of both low MACs and memory traffic. The new network achieves 35%, 36%, 30%, 32%, and 45% inference time reduction compared with FC-DenseNet-103, DenseNet-264, ResNet-50, ResNet-152, and SSD-VGG, respectively. We use tools including Nvidia profiler and ARM Scale-Sim to measure the memory traffic and verify that the inference latency is indeed proportional to the memory traffic consumption and the proposed network consumes low memory traffic. We conclude that one should take memory traffic into consideration when designing neural network architectures for high-resolution applications at the edge.

In this paper, we revive the use of old-fashioned handcrafted video representations for action recognition and put new life into these techniques via a CNN-based hallucination step. Despite of the use of RGB and optical flow frames, the I3D model (amongst others) thrives on combining its output with the Improved Dense Trajectory (IDT) and extracted with its low-level video descriptors encoded via Bag-of-Words (BoW) and Fisher Vectors (FV). Such a fusion of CNNs and handcrafted representations is time-consuming due to pre-processing, descriptor extraction, encoding and tuning parameters. Thus, we propose an end-to-end trainable network with streams which learn the IDT-based BoW/FV representations at the training stage and are simple to integrate with the I3D model. Specifically, each stream takes I3D feature maps ahead of the last 1D conv. layer and learns to `translate' these maps to BoW/FV representations. Thus, our model can hallucinate and use such synthesized BoW/FV representations at the testing stage. We show that even features of the entire I3D optical flow stream can be hallucinated thus simplifying the pipeline. Our model saves 20-55h of computations and yields state-of-the-art results on four publicly available datasets.

Estimating 3D human pose from a single image is a challenging task. This work attempts to address the uncertainty of lifting the detected 2D joints to the 3D space by introducing an intermediate state - Part-Centric Heatmap Triplets (HEMlets), which shortens the gap between the 2D observation and the 3D interpretation. The HEMlets utilize three joint-heatmaps to represent the relative depth information of the end-joints for each skeletal body part. In our approach, a Convolutional Network(ConvNet) is first trained to predict HEMlests from the input image, followed by a volumetric joint-heatmap regression. We leverage on the integral operation to extract the joint locations from the volumetric heatmaps, guaranteeing end-to-end learning. Despite the simplicity of the network design, the quantitative comparisons show a significant performance improvement over the best-of-grade method (by 20% on Human3.6M). The proposed method naturally supports training with "in-the-wild" images, where only weakly-annotated relative depth information of skeletal joints is available. This further improves the generalization ability of our model, as validated by qualitative comparisons on outdoor images.

MobileNets, a class of top-performing convolutional neural network architectures in terms of accuracy and efficiency trade-off, are increasingly used in many resource-aware vision applications. In this paper, we present Harmonious Bottleneck on two Orthogonal dimensions (HBO), a novel architecture unit, specially tailored to boost the accuracy of extremely lightweight MobileNets at the level of less than 40 MFLOPs. Unlike existing bottleneck designs that mainly focus on exploring the interdependencies among the channels of either groupwise or depthwise convolutional features, our HBO improves bottleneck representation while maintaining similar complexity via jointly encoding the feature interdependencies across both spatial and channel dimensions. It has two reciprocal components, namely spatial contraction-expansion and channel expansion-contraction, nested in a bilaterally symmetric structure. The combination of two interdependent transformations performing on orthogonal dimensions of feature maps enhances the representation and generalization ability of our proposed module, guaranteeing compelling performance with limited computational resource and power. By replacing the original bottlenecks in MobileNetV2 backbone with HBO modules, we construct HBONets which are evaluated on ImageNet classification, PASCAL VOC object detection and Market-1501 person re-identification. Extensive experiments show that with the severe constraint of computational budget our models outperform MobileNetV2 counterparts by remarkable margins of at most 6.6%, 6.3% and 5.0% on the above benchmarks respectively. Code and pretrained models are available at https://github.com/d-li14/HBONet.

Model size and inference speed/power have become a major challenge in the deployment of neural networks for many applications. A promising approach to address these problems is quantization. However, uniformly quantizing a model to ultra-low precision leads to significant accuracy degradation. A novel solution for this is to use mixed-precision quantization, as some parts of the network may allow lower precision as compared to other layers. However, there is no systematic way to determine the precision of different layers. A brute force approach is not feasible for deep networks, as the search space for mixed-precision is exponential in the number of layers. Another challenge is a similar factorial complexity for determining block-wise fine-tuning order when quantizing the model to a target precision. Here, we introduce Hessian AWare Quantization (HAWQ), a novel second-order quantization method to address these problems. HAWQ allows for the automatic selection of the relative quantization precision of each layer, based on the layer's Hessian spectrum. Moreover, HAWQ provides a deterministic fine-tuning order for quantizing layers. We show the results of our method on Cifar-10 using ResNet20, and on ImageNet using Inception-V3, ResNet50 and SqueezeNext models. Comparing HAWQ with state-of-the-art shows that we can achieve similar/better accuracy with 8x activation compression ratio on ResNet20, as compared to DNAS, and up to 1% higher accuracy with up to 14% smaller models on ResNet50 and Inception-V3, compared to recently proposed methods of RVQuant and HAQ. Furthermore, we show that we can quantize SqueezeNext to just 1MB model size while achieving above 68% top1 accuracy on ImageNet.

This paper presents a new large-scale dataset for recognition and temporal localization of human actions collected from Web videos. We refer to it as HACS (Human Action Clips and Segments). We leverage consensus and disagreement among visual classifiers to automatically mine candidate short clips from unlabeled videos, which are subsequently validated by human annotators. The resulting dataset is dubbed HACS Clips. Through a separate process we also collect annotations defining action segment boundaries. This resulting dataset is called HACS Segments. Overall, HACS Clips consists of 1.5M annotated clips sampled from 504K untrimmed videos, and HACS Segments contains 139K action segments densely annotated in 50K untrimmed videos spanning 200 action categories. HACS Clips contains more labeled examples than any existing video benchmark. This renders our dataset both a large-scale action recognition benchmark and an excellent source for spatiotemporal feature learning. In our transfer learning experiments on three target datasets, HACS Clips outperforms Kinetics-600, Moments-In-Time and Sports1M as a pretraining source. On HACS Segments, we evaluate state-of-the-art methods of action proposal generation and action localization, and highlight the new challenges posed by our dense temporal annotations.

Guided super-resolution is a unifying framework for several computer vision tasks where the inputs are a low-resolution source image of some target quantity (e.g., perspective depth acquired with a time-of-flight camera) and a high-resolution guide image from a different domain (e.g., a grey-scale image from a conventional camera); and the target output is a high-resolution version of the source (in our example, a high-res depth map). The standard way of looking at this problem is to formulate it as a super-resolution task, i.e., the source image is upsampled to the target resolution, while transferring the missing high-frequency details from the guide. Here, we propose to turn that interpretation on its head and instead see it as a pixel-to-pixel mapping of the guide image to the domain of the source image. The pixel-wise mapping is parametrised as a multi-layer perceptron, whose weights are learned by minimising the discrepancies between the source image and the downsampled target image. Importantly, our formulation makes it possible to regularise only the mapping function, while avoiding regularisation of the outputs; thus producing crisp, natural-looking images. The proposed method is unsupervised, using only the specific source and guide images to fit the mapping. We evaluate our method on two different tasks, super-resolution of depth maps and of tree height maps. In both cases, we clearly outperform recent baselines in quantitative comparisons, while delivering visually much sharper outputs.