CVPR 2018 论文列表

Human actions in videos are three-dimensional (3D) signals. Recent attempts use 3D convolutional neural networks (CNNs) to explore spatio-temporal information for human action recognition. Though promising, 3D CNNs have not achieved high performanceon on this task with respect to their well-established two-dimensional (2D) counterparts for visual recognition in still images. We argue that the high training complexity of spatio-temporal fusion and the huge memory cost of 3D convolution hinder current 3D CNNs, which stack 3D convolutions layer by layer, by outputting deeper feature maps that are crucial for high-level tasks. We thus propose a Mixed Convolutional Tube (MiCT) that integrates 2D CNNs with the 3D convolution module to generate deeper and more informative feature maps, while reducing training complexity in each round of spatio-temporal fusion. A new end-to-end trainable deep 3D network, MiCT-Net, is also proposed based on the MiCT to better explore spatio-temporal information in human actions. Evaluations on three well-known benchmark datasets (UCF101, Sport-1M and HMDB-51) show that the proposed MiCT-Net significantly outperforms the original 3D CNNs. Compared with state-of-the-art approaches for action recognition on UCF101 and HMDB51, our MiCT-Net yields the best performance.

We present a learning-based approach for synthesizing facial geometry at medium and fine scales from diffusely-lit facial texture maps. When applied to an image sequence, the synthesized detail is temporally coherent. Unlike current state-of-the-art methods, which assume "dark is deep", our model is trained with measured facial detail collected using polarized gradient illumination in a Light Stage. This enables us to produce plausible facial detail across the entire face, including where previous approaches may incorrectly interpret dark features as concavities such as at moles, hair stubble, and occluded pores. Instead of directly inferring 3D geometry, we propose to encode fine details in high-resolution displacement maps which are learned through a hybrid network adopting the state-of-the-art image-to-image translation network and super resolution network. To effectively capture geometric detail at both mid- and high frequencies, we factorize the learning into two separate sub-networks, enabling the full range of facial detail to be modeled. Results from our learning-based approach compare favorably with a high-quality active facial scanning technique, and require only a single passive lighting condition without a complex scanning setup.

In this paper, we introduce the new ideas of augmenting Convolutional Neural Networks (CNNs) with Memory and learning to learn the network parameters for the unlabelled images on the fly in one-shot learning. Specifically, we present Memory Matching Networks (MM-Net) --- a novel deep architecture that explores the training procedure, following the philosophy that training and test conditions must match. Technically, MM-Net writes the features of a set of labelled images (support set) into memory and reads from memory when performing inference to holistically leverage the knowledge in the set. Meanwhile, a Contextual Learner employs the memory slots in a sequential manner to predict the parameters of CNNs for unlabelled images. The whole architecture is trained by once showing only a few examples per class and switching the learning from minibatch to minibatch, which is tailored for one-shot learning when presented with a few examples of new categories at test time. Unlike the conventional one-shot learning approaches, our MM-Net could output one unified model irrespective of the number of shots and categories. Extensive experiments are conducted on two public datasets, i.e., Omniglot and emph{mini}ImageNet, and superior results are reported when compared to state-of-the-art approaches. More remarkably, our MM-Net improves one-shot accuracy on Omniglot from 98.95% to 99.28% and from 49.21% to 53.37% on emph{mini}ImageNet.

We present a novel online unsupervised method for face identity learning from video streams. The method exploits deep face descriptors together with a memory based learning mechanism that takes advantage of the temporal coherence of visual data. Specifically, we introduce a discriminative descriptor matching solution based on Reverse Nearest Neighbour and a forgetting strategy that detect redundant descriptors and discard them appropriately while time progresses. It is shown that the proposed learning procedure is asymptotically stable and can be effectively used in relevant applications like multiple face identification and tracking from unconstrained video streams. Experimental results show that the proposed method achieves comparable results in the task of multiple face tracking and better performance in face identification with offline approaches exploiting future information. Code will be publicly available.

Single-view depth prediction is a fundamental problem in computer vision. Recently, deep learning methods have led to significant progress, but such methods are limited by the available training data. Current datasets based on 3D sensors have key limitations, including indoor-only images (NYU), small numbers of training examples (Make3D), and sparse sampling (KITTI). We propose to use multi-view Internet photo collections, a virtually unlimited data source, to generate training data via modern structure-from-motion and multi-view stereo (MVS) methods, and present a large depth dataset called MegaDepth based on this idea. Data derived from MVS comes with its own challenges, including noise and unreconstructable objects. We address these challenges with new data cleaning methods, as well as automatically augmenting our data with ordinal depth relations generated using semantic segmentation. We validate the use of large amounts of Internet data by showing that models trained on MegaDepth exhibit strong generalization—not only to novel scenes, but also to other diverse datasets including Make3D, KITTI, and DIW, even when no images from those datasets are seen during training.

The development of object detection in the era of deep learning, from R-CNN [11], Fast/Faster R-CNN [10, 31] to recent Mask R-CNN [14] and RetinaNet [24], mainly come from novel network, new framework, or loss design. How- ever, mini-batch size, a key factor for the training of deep neural networks, has not been well studied for object detec- tion. In this paper, we propose a Large Mini-Batch Object Detector (MegDet) to enable the training with a large mini- batch size up to 256, so that we can effectively utilize at most 128 GPUs to significantly shorten the training time. Technically, we suggest a warmup learning rate policy and Cross-GPU Batch Normalization, which together allow us to successfully train a large mini-batch detector in much less time (e.g., from 33 hours to 4 hours), and achieve even better accuracy. The MegDet is the backbone of our sub- mission (mmAP 52.5%) to COCO 2017 Challenge, where we won the 1st place of Detection task.

Age estimation has broad application prospects of many fields, such as video surveillance, social networking, and human-computer interaction. However, many of the published age estimation approaches simply treat the age estimation as an exact age regression problem, and thus did not leverage a distribution's robustness in representing labels with ambiguity such as ages. In this paper, we propose a new loss function, called mean-variance loss, for robust age estimation via distribution learning. Specifically, the mean-variance loss consists of a mean loss, which penalizes difference between the mean of the estimated age distribution and the ground-truth age, and a variance loss, which penalizes the variance of the estimated age distribution to ensure a concentrated distribution. The proposed mean-variance loss and softmax loss are embedded jointly into Convolutional Neural Networks (CNNs) for age estimation, and the network weights are optimized via stochastic gradient descent (SGD) in an end-to-end learning way. Experimental results on a number of challenging face aging databases (FG-NET, MORPH Album II, and CLAP2016) show that the proposed approach outperforms the state-of-the-art methods by a large margin using a single model.

In this work, we present a method for unsupervised domain adaptation. Many adversarial learning methods train domain classifier networks to distinguish the features as either a source or target and train a feature generator network to mimic the discriminator. Two problems exist with these methods. First, the domain classifier only tries to distinguish the features as a source or target and thus does not consider task-specific decision boundaries between classes. Therefore, a trained generator can generate ambiguous features near class boundaries. Second, these methods aim to completely match the feature distributions between different domains, which is difficult because of each domain's characteristics. To solve these problems, we introduce a new approach that attempts to align distributions of source and target by utilizing the task-specific decision boundaries. We propose to maximize the discrepancy between two classifiers' outputs to detect target samples that are far from the support of the source. A feature generator learns to generate target features near the support to minimize the discrepancy. Our method outperforms other methods on several datasets of image classification and semantic segmentation. The codes are available at url{https://github.com/mil-tokyo/MCD_DA}

In this paper, we develop novel, efficient 2D encodings for 3D geometry, which enable reconstructing full 3D shapes from a single image at high resolution. The key idea is to pose 3D shape reconstruction as a 2D prediction problem. To that end, we first develop a simple baseline network that predicts entire voxel tubes at each pixel of a reference view. By leveraging well-proven architectures for 2D pixel-prediction tasks, we attain state-of-the-art results, clearly outperforming purely voxel-based approaches. We scale this baseline to higher resolutions by proposing a memory-efficient shape encoding, which recursively decomposes a 3D shape into nested shape layers, similar to the pieces of a Matryoshka doll. This allows reconstructing highly detailed shapes with complex topology, as demonstrated in extensive experiments; we clearly outperform previous octree-based approaches despite having a much simpler architecture using standard network components. Our Matryoshka networks further enable reconstructing shapes from IDs or shape similarity, as well as shape sampling.

We propose a new error measure for matching pixels that is based on co-occurrence statistics. The measure relies on a co-occurrence matrix that counts the number of times pairs of pixel values co-occur within a window. The error incurred by matching a pair of pixels is inverse proportional to the probability that their values co-occur together, and not their color difference. This measure also works with features other than color, e.g. deep features. We show that this improves the state-of-the-art performance of template matching on standard benchmarks. We then propose an embedding scheme that maps the input image to an embedded image such that the Euclidean distance between pixel values in the embedded space resembles the co-occurrence statistics in the original space. This lets us run existing vision algorithms on the embedded images and enjoy the power of co-occurrence statistics for free. We demonstrate this on two algorithms, the Lucas-Kanade image registration and the Kernelized Correlation Filter (KCF) tracker. Experiments show that performance of each algorithm improves by about 10%.

Generative Adversarial Nets (GANs) and Conditonal GANs (CGANs) show that using a trained network as loss function (discriminator) enables to synthesize highly structured outputs (e.g. natural images). However, applying a discriminator network as a universal loss function for common supervised tasks (e.g. semantic segmentation, line detection, depth estimation) is considerably less successful. We argue that the main difficulty of applying CGANs to supervised tasks is that the generator training consists of optimizing a loss function that does not depend directly on the ground truth labels. To overcome this, we propose to replace the discriminator with a matching network taking into account both the ground truth outputs as well as the generated examples. As a consequence, the generator loss function also depends on the targets of the training examples, thus facilitating learning. We demonstrate on three computer vision tasks that this approach can significantly outperform CGANs achieving comparable or superior results to task-specific solutions and results in stable training. Importantly, this is a general approach that does not require the use of task-specific loss functions.

In this work, we tackle the problem of instance segmentation, the task of simultaneously solving object detection and semantic segmentation. Towards this goal, we present a model, called MaskLab, which produces three outputs: box detection, semantic segmentation, and direction prediction. Building on top of the Faster-RCNN object detector, the predicted boxes provide accurate localization of object instances. Within each region of interest, MaskLab performs foreground/background segmentation by combining semantic and direction prediction. Semantic segmentation assists the model in distinguishing between objects of different semantic classes including background, while the direction prediction, estimating each pixel's direction towards its corresponding center, allows separating instances of the same semantic class. Moreover, we explore the effect of incorporating recent successful methods from both segmentation and detection (eg, atrous convolution and hypercolumn). Our proposed model is evaluated on the COCO instance segmentation benchmark and shows comparable performance with other state-of-art models.

Person Re-identification (ReID) is an important yet challenging task in computer vision. Due to the diverse background clutters, variations on viewpoints and body poses, it is far from solved. How to extract discriminative and robust features invariant to background clutters is the core problem. In this paper, we first introduce the binary segmentation masks to construct synthetic RGB-Mask pairs as inputs, then we design a mask-guided contrastive attention model (MGCAM) to learn features separately from the body and background regions. Moreover, we propose a novel region-level triplet loss to restrain the features learnt from different regions, i.e., pulling the features from the full image and body region close, whereas pushing the features from backgrounds away. We may be the first one to successfully introduce the binary mask into person ReID task and the first one to propose region-level contrastive learning. We evaluate the proposed method on three public datasets, including MARS, Market-1501 and CUHK03. Extensive experimental results show that the proposed method is effective and achieves the state-of-the-art results. Mask and code will be released upon request.

Autonomous agents need to reason about the world beyond their instantaneous sensory input. Integrating information over time, however, requires switching from an egocentric representation of a scene to an allocentric one, expressed in the world reference frame. It must also be possible to update the representation dynamically, which requires localizing and registering the sensor with respect to it. In this paper, we develop a differentiable module that satisfies such requirements, while being robust, efficient, and suitable for integration in end-to-end deep networks. The module contains an allocentric spatial memory that can be accessed associatively by feeding to it the current sensory input, resulting in localization, and then updated using an LSTM or similar mechanism. We formulate efficient localization and registration of sensory information as a dual pair of convolution/deconvolution operators in memory space. The map itself is a 2.5D representation of an environment storing information that a deep neural network module learns to distill from RGBD input. The result is a map that contains multi-task information, different from classical approaches to mapping such as structure-from-motion. We present results using synthetic mazes, a dataset of hours of recorded gameplay of the classic game Doom, and the very recent Active Vision Dataset of real images captured from a robot.

When learning 3D shapes we are usually interested in their intrinsic geometry rather than in their orientation. To deal with the orientation variations the usual trick consists in augmenting the data to exhibit all possible variability, and thus let the model learn both the geometry as well as the rotations. In this paper we introduce a new autoencoder model for encoding and synthesis of 3D shapes. To get rid of undesirable input variability our model learns a manifold in a quotient space of the input space. Typically, we propose to quotient the space of 3D models by the action of rotations. Thus, our quotient autoencoder allows to directly learn in the space of interest, ignoring side information. This is reflected in better performances on reconstruction and interpolation tasks, as our experiments show that our model outperforms a vanilla autoencoder on the well-known Shapenet dataset. Moreover, our model learns a rotation-invariant representation, leading to interesting results in shapes co-alignment. Finally, we extend our quotient autoencoder to quotient by non-rigid transformations.

Recurrent neural networks (RNNs) have emerged as a powerful model for a broad range of machine learning problems that involve sequential data. While an abundance of work exists to understand and improve RNNs in the context of language and audio signals such as language modeling and speech recognition, relatively little attention has been paid to analyze or modify RNNs for visual sequences, which by nature have distinct properties. In this paper, we aim to bridge this gap and present the first large-scale exploration of RNNs for visual sequence learning. In particular, with the intention of leveraging the strong generalization capacity of pre-trained convolutional neural networks (CNNs), we propose a novel and effective approach, PreRNN, to make pre-trained CNNs recurrent by transforming convolutional layers or fully connected layers into recurrent layers. We conduct extensive evaluations on three representative visual sequence learning tasks: sequential face alignment, dynamic hand gesture recognition, and action recognition. Our experiments reveal that PreRNN consistently outperforms the traditional RNNs and achieves state-of-the-art results on the three applications, suggesting that PreRNN is more suitable for visual sequence learning.

Recent approaches on trajectory forecasting use tracklets to predict the future positions of pedestrians exploiting Long Short Term Memory (LSTM) architectures. This paper shows that adding vislets, that is, short sequences of head pose estimations, allows to increase significantly the trajectory forecasting performance. We then propose to use vislets in a novel framework called MX-LSTM, capturing the interplay between tracklets and vislets thanks to a joint unconstrained optimization of full covariance matrices during the LSTM backpropagation. At the same time, MX-LSTM predicts the future head poses, increasing the standard capabilities of the long-term trajectory forecasting approaches. With standard head pose estimators and an attentional-based social pooling, Mixing-LSTM scores the new trajectory forecasting state-of-the-art in all the considered datasets (Zara01, Zara02, UCY, and TownCentre) with a dramatic margin when the pedestrians slow down, a case where most of the forecasting approaches struggle to provide an accurate solution.

In this paper, we address referring expression comprehension: localizing an image region described by a natural language expression. While most recent work treats expressions as a single unit, we propose to decompose them into three modular components related to subject appearance, location, and relationship to other objects. This allows us to flexibly adapt to expressions containing different types of information in an end-to-end framework. In our model, which we call the Modular Attention Network (MAttNet), two types of attention are utilized: language-based attention that learns the module weights as well as the word/phrase attention that each module should focus on; and visual attention that allows the subject and relationship modules to focus on relevant image components. Module weights combine scores from all three modules dynamically to output an overall score. Experiments show that MAttNet outperforms previous state-of-the-art methods by a large margin on both bounding-box-level and pixel-level comprehension tasks.

Video captioning which automatically translates video clips into natural language sentences is a very important task in computer vision. By virtue of recent deep learning technologies, video captioning has made great progress. However, learning an effective mapping from the visual sequence space to the language space is still a challenging problem due to the long-term multimodal dependency modelling and semantic misalignment. Inspired by the facts that memory modelling poses potential advantages to long-term sequential problems [35] and working memory is the key factor of visual attention [33], we propose a Multimodal Memory Model (M3) to describe videos, which builds a visual and textual shared memory to model the long-term visual-textual dependency and further guide visual attention on described visual targets to solve visual-textual alignments. Specifically, similar to [10], the proposed M3 attaches an external memory to store and retrieve both visual and textual contents by interacting with video and sentence with multiple read and write operations. To evaluate the proposed model, we perform experiments on two public datasets: MSVD and MSR-VTT. The experimental results demonstrate that our method outperforms most of the state-of-the-art methods in terms of BLEU and METEOR.