ICCV 2019 论文列表

Action recognition has seen a dramatic performance improvement in the last few years. Most of the current state-of-the-art literature either aims at improving performance through changes to the backbone CNN network, or exploring different trade-offs between computational efficiency and performance, again through altering the backbone network. However, almost all of these works maintain the same last layers of the network, which simply consist of a global average pooling followed by a fully connected layer. In this work we focus on how to improve the representation capacity of the network, but rather than altering the backbone, we focus on improving the last layers of the network, where changes have low impact in terms of computational cost. In particular, we hypothesize that current architectures have poor sensitivity to finer details and we exploit recent advances in the fine-grained recognition literature to improve our model in this aspect. With the proposed approach, we obtain state-of-the-art performance on Kinetics-400 and Something-Something-V1, the two major large-scale action recognition benchmarks.

In this paper, we propose a monocular 3D object detection framework in the domain of autonomous driving. Unlike previous image-based methods which focus on RGB feature extracted from 2D images, our method solves this problem in the reconstructed 3D space in order to exploit 3D contexts explicitly. To this end, we first leverage a stand-alone module to transform the input data from 2D image plane to 3D point clouds space for a better input representation, then we perform the 3D detection using PointNet backbone net to obtain objects' 3D locations, dimensions and orientations. To enhance the discriminative capability of point clouds, we propose a multi-modal feature fusion module to embed the complementary RGB cue into the generated point clouds representation. We argue that it is more effective to infer the 3D bounding boxes from the generated 3D scene space (i.e., X,Y, Z space) compared to the image plane (i.e., R,G,B image plane). Evaluation on the challenging KITTI dataset shows that our approach boosts the performance of state-of-the-art monocular approach by a large margin.

This work describes Barnes-Hut Rigid Gravitational Approach (BH-RGA) -- a new rigid point set registration method relying on principles of particle dynamics. Interpreting the inputs as two interacting particle swarms, we directly minimise the gravitational potential energy of the system using non-linear least squares. Compared to solutions obtained by solving systems of second-order ordinary differential equations, our approach is more robust and less dependent on the parameter choice. We accelerate otherwise exhaustive particle interactions with a Barnes-Hut tree and efficiently handle massive point sets in quasilinear time while preserving the globally multiply-linked character of interactions. Among the advantages of BH-RGA is the possibility to define boundary conditions or additional alignment cues through varying point masses. Systematic experiments demonstrate that BH-RGA surpasses performances of baseline methods in terms of the convergence basin and accuracy when handling incomplete, noisy and perturbed data. The proposed approach also positively compares to the competing method for the alignment with prior matches.

Recent years have witnessed the success of learning to hash in fast large-scale image retrieval. As deep learning has shown its superior performance on many computer vision applications, recent designs of learning-based hashing models have been moving from shallow ones to deep architectures. However, based on our analysis, we find that gradient descent based algorithms used in deep hashing models would potentially cause hash codes of a pair of training instances to be updated towards the directions of each other simultaneously during optimization. In the worst case, the paired hash codes switch their directions after update, and consequently, their corresponding distance in the Hamming space remain unchanged. This makes the overall learning process highly inefficient. To address this issue, we propose a new deep hashing model integrated with a novel gradient attention mechanism. Extensive experimental results on three benchmark datasets show that our proposed algorithm is able to accelerate the learning process and obtain competitive retrieval performance compared with state-of-the-art deep hashing models.

Channel Pruning, widely used for accelerating Convolutional Neural Networks, is an NP-hard problem due to the inter-layer dependency of channel redundancy. Existing methods generally ignored the above dependency for computation simplicity. To solve the problem, under the Bayesian framework, we here propose a layer-wise Recursive Bayesian Pruning method (RBP). A new dropout-based measurement of redundancy, which facilitate the computation of posterior assuming inter-layer dependency, is introduced. Specifically, we model the noise across layers as a Markov chain and target its posterior to reflect the inter-layer dependency. Considering the closed form solution for posterior is intractable, we derive a sparsity-inducing Dirac-like prior which regularizes the distribution of the designed noise to automatically approximate the posterior. Compared with the existing methods, no additional overhead is required when the inter-layer dependency assumed. The redundant channels can be simply identified by tiny dropout noise and directly pruned layer by layer. Experiments on popular CNN architectures have shown that the proposed method outperforms several state-of-the-arts. Particularly, we achieve up to 5.0x, 2.2x and 1.7x FLOPs reduction with little accuracy loss on the large scale dataset ILSVRC2012 for VGG16, ResNet50 and MobileNetV2, respectively.

We propose an Adaptive Weighted Spatiotemporal Distillation (AWSD) technique for video representation by encoding the appearance and dynamics of the videos into a single RGB image map. This is obtained by adaptively dividing the videos into small segments and comparing two consecutive segments. This allows using pre-trained models on still images for video classification while successfully capturing the spatiotemporal variations in the videos. The adaptive segment selection enables effective encoding of the essential discriminative information of untrimmed videos. Based on Gaussian Scale Mixture, we compute the weights by extracting the mutual information between two consecutive segments. Unlike pooling-based methods, our AWSD gives more importance to the frames that characterize actions or events thanks to its adaptive segment length selection. We conducted extensive experimental analysis to evaluate the effectiveness of our proposed method and compared our results against those of recent state-of-the-art methods on four benchmark datatsets, including UCF101, HMDB51, ActivityNet v1.3, and Maryland. The obtained results on these benchmark datatsets showed that our method significantly outperforms earlier works and sets the new state-of-the-art performance in video classification. Code is available at the project webpage: https://mohammadt68.github.io/AWSD/

The learning of Transformation-Equivariant Representations (TERs), which is introduced by Hinton et al. [??], has been considered as a principle to reveal visual structures under various transformations. It contains the celebrated Convolutional Neural Networks (CNNs) as a special case that only equivary to the translations. In contrast, we seek to train TERs for a generic class of transformations and train them in an unsupervised fashion. To this end, we present a novel principled method by Autoencoding Variational Transformations (AVT), compared with the conventional approach to autoencoding data. Formally, given transformed images, the AVT seeks to train the networks by maximizing the mutual information between the transformations and representations. This ensures the resultant TERs of individual images contain the intrinsic information about their visual structures that would equivary extricably under various transformations in a generalized nonlinear case. Technically, we show that the resultant optimization problem can be efficiently solved by maximizing a variational lower-bound of the mutual information. This variational approach introduces a transformation decoder to approximate the intractable posterior of transformations, resulting in an autoencoding architecture with a pair of the representation encoder and the transformation decoder. Experiments demonstrate the proposed AVT model sets a new record for the performances on unsupervised tasks, greatly closing the performance gap to the supervised models.

In this paper we propose an attentive recurrent generative adversarial network (ARGAN) to detect and remove shadows in an image. The generator consists of multiple progressive steps. At each step a shadow attention detector is firstly exploited to generate an attention map which specifies shadow regions in the input image. Given the attention map, a negative residual by a shadow remover encoder will recover a shadow-lighter or even a shadow-free image. The discriminator is designed to classify whether the output image in the last progressive step is real or fake. Moreover, ARGAN is suitable to be trained with a semi-supervised strategy to make full use of sufficient unsupervised data. The experiments on four public datasets have demonstrated that our ARGAN is robust to detect both simple and complex shadows and to produce more realistic shadow removal results. It outperforms the state-of-the-art methods, especially in detail of recovering shadow areas.

Deep learning has thrived by training on large-scale datasets. However, in robotics applications sample efficiency is critical. We propose a novel adaptive masked proxies method that constructs the final segmentation layer weights from few labelled samples. It utilizes multi-resolution average pooling on base embeddings masked with the label to act as a positive proxy for the new class, while fusing it with the previously learned class signatures. Our method is evaluated on PASCAL-5^i dataset and outperforms the state-of-the-art in the few-shot semantic segmentation. Unlike previous methods, our approach does not require a second branch to estimate parameters or prototypes, which enables it to be used with 2-stream motion and appearance based segmentation networks. We further propose a novel setup for evaluating continual learning of object segmentation which we name incremental PASCAL (iPASCAL) where our method outperforms the baseline method. Our code is publicly available at https://github.com/MSiam/AdaptiveMaskedProxies.

Large datasets are the cornerstone of recent advances in computer vision using deep learning. In contrast, existing human motion capture (mocap) datasets are small and the motions limited, hampering progress on learning models of human motion. While there are many different datasets available, they each use a different parameterization of the body, making it difficult to integrate them into a single meta dataset. To address this, we introduce AMASS, a large and varied database of human motion that unifies 15 different optical marker-based mocap datasets by representing them within a common framework and parameterization. We achieve this using a new method, MoSh++, that converts mocap data into realistic 3D human meshes represented by a rigged body model. Here we use SMPL [Loper et al., 2015], which is widely used and provides a standard skeletal representation as well as a fully rigged surface mesh. The method works for arbitrary marker sets, while recovering soft-tissue dynamics and realistic hand motion. We evaluate MoSh++ and tune its hyperparameters using a new dataset of 4D body scans that are jointly recorded with markerbased mocap. The consistent representation of AMASS makes it readily useful for animation, visualization, and generating training data for deep learning. Our dataset is significantly richer than previous human motion collections, having more than 40 hours of motion data, spanning over 300 subjects, more than 11000 motions, and will be publicly available to the research community.

Designing an effective loss function plays an important role in visual analysis. Most existing loss function designs rely on hand-crafted heuristics that require domain experts to explore the large design space, which is usually sub-optimal and time-consuming. In this paper, we propose AutoML for Loss Function Search (AM-LFS) which leverages REINFORCE to search loss functions during the training process. The key contribution of this work is the design of search space which can guarantee the generalization and transferability on different vision tasks by including a bunch of existing prevailing loss functions in a unified formulation. We also propose an efficient optimization framework which can dynamically optimize the parameters of loss function's distribution during training. Extensive experimental results on four benchmark datasets show that, without any tricks, our method outperforms existing hand-crafted loss functions in various computer vision tasks.

Most video object segmentation approaches process objects separately. This incurs high computational cost when multiple objects exist. In this paper, we propose AGSS-VOS to segment multiple objects in one feed-forward path via instance-agnostic and instance-specific modules. Information from the two modules is fused via an attention-guided decoder to simultaneously segment all object instances in one path. The whole framework is end-to-end trainable with instance IoU loss. Experimental results on Youtube- VOS and DAVIS-2017 dataset demonstrate that AGSS-VOS achieves competitive results in terms of both accuracy and efficiency.

As designing appropriate Convolutional Neural Network (CNN) architecture in the context of a given application usually involves heavy human works or numerous GPU hours, the research community is soliciting the architecture-neutral CNN structures, which can be easily plugged into multiple mature architectures to improve the performance on our real-world applications. We propose Asymmetric Convolution Block (ACB), an architecture-neutral structure as a CNN building block, which uses 1D asymmetric convolutions to strengthen the square convolution kernels. For an off-the-shelf architecture, we replace the standard square-kernel convolutional layers with ACBs to construct an Asymmetric Convolutional Network (ACNet), which can be trained to reach a higher level of accuracy. After training, we equivalently convert the ACNet into the same original architecture, thus requiring no extra computations anymore. We have observed that ACNet can improve the performance of various models on CIFAR and ImageNet by a clear margin. Through further experiments, we attribute the effectiveness of ACB to its capability of enhancing the model's robustness to rotational distortions and strengthening the central skeleton parts of square convolution kernels.

Image and sentence matching has drawn much attention recently, but due to the lack of sufficient pairwise data for training, most previous methods still cannot well associate those challenging pairs of images and sentences containing rarely appeared regions and words, i.e., few-shot content. In this work, we study this challenging scenario as few-shot image and sentence matching, and accordingly propose an Aligned Cross-Modal Memory (ACMM) model to memorize the rarely appeared content. Given a pair of image and sentence, the model first includes an aligned memory controller network to produce two sets of semantically-comparable interface vectors through cross-modal alignment. Then the interface vectors are used by modality-specific read and update operations to alternatively interact with shared memory items. The memory items persistently memorize cross-modal shared semantic representations, which can be addressed out to better enhance the representation of few-shot content. We apply the proposed model to both conventional and few-shot image and sentence matching tasks, and demonstrate its effectiveness by achieving the state-of-the-art performance on two benchmark datasets.

Recent works have made great progress in semantic segmentation by exploiting richer context, most of which are designed from a spatial perspective. In contrast to previous works, we present the concept of class center which extracts the global context from a categorical perspective. This class-level context describes the overall representation of each class in an image. We further propose a novel module, named Attentional Class Feature (ACF) module, to calculate and adaptively combine different class centers according to each pixel. Based on the ACF module, we introduce a coarse-to-fine segmentation network, called Attentional Class Feature Network (ACFNet), which can be composed of an ACF module and any off-the-shell segmentation network (base network). In this paper, we use two types of base networks to evaluate the effectiveness of ACFNet. We achieve new state-of-the-art performance of 81.85% mIoU on Cityscapes dataset with only finely annotated data used for training.

Deep neural networks exhibit exceptional accuracy when they are trained and tested on the same data distributions. However, neural classifiers are often extremely brittle when confronted with domain shift---changes in the input distribution that occur over time. We present ACE, a framework for semantic segmentation that dynamically adapts to changing environments over time. By aligning the distribution of labeled training data from the original source domain with the distribution of incoming data in a shifted domain, ACE synthesizes labeled training data for environments as it sees them. This stylized data is then used to update a segmentation model so that it performs well in new environments. To avoid forgetting knowledge from past environments, we introduce a memory that stores feature statistics from previously seen domains. These statistics can be used to replay images in any of the previously observed domains, thus preventing catastrophic forgetting. In addition to standard batch training using stochastic gradient decent (SGD), we also experiment with fast adaptation methods based on adaptive meta-learning. Extensive experiments are conducted on two datasets from SYNTHIA, the results demonstrate the effectiveness of the proposed approach when adapting to a number of tasks.

Attention mechanisms have been found effective for person re-identification (Re-ID). However, the learned "attentive" features are often not naturally uncorrelated or "diverse", which compromises the retrieval performance based on the Euclidean distance. We advocate the complementary powers of attention and diversity for Re-ID, by proposing an Attentive but Diverse Network (ABD-Net). ABD-Net seamlessly integrates attention modules and diversity regularizations throughout the entire network to learn features that are representative, robust, and more discriminative. Specifically, we introduce a pair of complementary attention modules, focusing on channel aggregation and position awareness, respectively. Then, we plug in a novel orthogonality constraint that efficiently enforces diversity on both hidden activations and weights. Through an extensive set of ablation study, we verify that the attentive and diverse terms each contributes to the performance boosts of ABD-Net. It consistently outperforms existing state-of-the-art methods on there popular person Re-ID benchmarks.

For 3D hand and body pose estimation task in depth image, a novel anchor-based approach termed Anchor-to-Joint regression network (A2J) with the end-to-end learning ability is proposed. Within A2J, anchor points able to capture global-local spatial context information are densely set on depth image as local regressors for the joints. They contribute to predict the positions of the joints in ensemble way to enhance generalization ability. The proposed 3D articulated pose estimation paradigm is different from the state-of-the-art encoder-decoder based FCN, 3D CNN and point-set based manners. To discover informative anchor points towards certain joint, anchor proposal procedure is also proposed for A2J. Meanwhile 2D CNN (i.e., ResNet- 50) is used as backbone network to drive A2J, without using time-consuming 3D convolutional or deconvolutional layers. The experiments on 3 hand datasets and 2 body datasets verify A2J's superiority. Meanwhile, A2J is of high running speed around 100 FPS on single NVIDIA 1080Ti GPU.

Objects are usually organized in a hierarchical structure in which each coarse category (e.g., big cat) corresponds to a superclass of several fine categories (e.g., cheetah, leopard). The objects grouped within the same coarse category, but in different fine categories, usually share a set of global visual features; however, these objects have distinctive local properties that characterize them at a fine level. This paper addresses the challenge of fine image classification in a weakly supervised fashion, whereby a subset of images is tagged by fine labels, while the remaining are tagged by coarse labels. We propose a new deep model that leverages coarse images to improve the classification performance of fine images within the coarse category. Our model is an end to end framework consisting of a Convolutional Neural Network (CNN) which uses both fine and coarse images to tune its parameters. The CNN outputs are then fanned out into two separate branches such that the first branch uses a supervised low rank self expressive layer to project the CNN outputs to the low rank subspaces to capture the global structures for the coarse classification, while the other branch uses a supervised sparse self expressive layer to project them to the sparse subspaces to capture the local structures for the fine classification. Our deep model uses coarse images in conjunction with fine images to jointly explore the low rank and sparse subspaces by sharing the parameters during the training which causes the data points obtained by the CNN to be well-projected to both sparse and low rank subspaces for classification.