ICCV 2019 Papers

Existing video-based human pose estimation methods extensively apply large networks onto every frame in the video to localize body joints, which suffer high computational cost and hardly meet the low-latency requirement in realistic applications. To address this issue, we propose a novel Dynamic Kernel Distillation (DKD) model to facilitate small networks for estimating human poses in videos, thus significantly lifting the efficiency. In particular, DKD introduces a light-weight distillator to online distill pose kernels via leveraging temporal cues from the previous frame in a one-shot feed-forward manner. Then, DKD simplifies body joint localization into a matching procedure between the pose kernels and the current frame, which can be efficiently computed via simple convolution. In this way, DKD fast transfers pose knowledge from one frame to provide compact guidance for body joint localization in the following frame, which enables utilization of small networks in video-based pose estimation. To facilitate the training process, DKD exploits a temporally adversarial training strategy that introduces a temporal discriminator to help generate temporally coherent pose kernels and pose estimation results within a long range. Experiments on Penn Action and Sub-JHMDB benchmarks demonstrate outperforming efficiency of DKD, specifically, 10x flops reduction and 2x speedup over previous best model, and its state-of-the-art accuracy.

Human attribute analysis is a challenging task in the field of computer vision. One of the significant difficulties is brought from largely imbalance-distributed data. Conventional techniques such as re-sampling and cost-sensitive learning require prior-knowledge to train the system. To address this problem, we propose a unified framework called Dynamic Curriculum Learning (DCL) to adaptively adjust the sampling strategy and loss weight in each batch, which results in better ability of generalization and discrimination. Inspired by curriculum learning, DCL consists of two-level curriculum schedulers: (1) sampling scheduler which manages the data distribution not only from imbalance to balance but also from easy to hard; (2) loss scheduler which controls the learning importance between classification and metric learning loss. With these two schedulers, we achieve state-of-the-art performance on the widely used face attribute dataset CelebA and pedestrian attribute dataset RAP.

Establishing semantic correspondence is a core problem in computer vision and remains challenging due to large intra-class variations and lack of annotated data. In this paper, we aim to incorporate global semantic context in a flexible manner to overcome the limitations of prior work that relies on local semantic representations. To this end, we first propose a context-aware semantic representation that incorporates spatial layout for robust matching against local ambiguities. We then develop a novel dynamic fusion strategy based on attention mechanism to weave the advantages of both local and context features by integrating semantic cues from multiple scales. We instantiate our strategy by designing an end-to-end learnable deep network, named as Dynamic Context Correspondence Network (DCCNet). To train the network, we adopt a multi-auxiliary task loss to improve the efficiency of our weakly-supervised learning procedure. Our approach achieves superior or competitive performance over previous methods on several challenging datasets, including PF-Pascal, PF-Willow, and TSS, demonstrating its effectiveness and generality.

The design of anchors is critical to the performance of one-stage detectors. Recently, the anchor refinement module (ARM) has been proposed to adjust the initialization of default anchors, providing the detector a better anchor reference. However, this module brings another problem: all pixels at a feature map have the same receptive field while the anchors associated with each pixel have different positions and sizes. This discordance may lead to a less effective detector. In this paper, we present a dynamic feature selection operation to select new pixels in a feature map for each refined anchor received from the ARM. The pixels are selected based on the new anchor position and size so that the receptive filed of these pixels can fit the anchor areas well, which makes the detector, especially the regression part, much easier to optimize. Furthermore, to enhance the representation ability of selected feature pixels, we design a bidirectional feature fusion module by combining features from early and deep layers. Extensive experiments on both PASCAL VOC and COCO demonstrate the effectiveness of our dynamic anchor feature selection (DAFS) operation. For the case of high IoU threshold, our DAFS can improve the mAP by a large margin.

Recently, consistency-based methods have achieved state-of-the-art results in semi-supervised learning (SSL). These methods always involve two roles, an explicit or implicit teacher model and a student model, and penalize predictions under different perturbations by a consistency constraint. However, the weights of these two roles are tightly coupled since the teacher is essentially an exponential moving average (EMA) of the student. In this work, we show that the coupled EMA teacher causes a performance bottleneck. To address this problem, we introduce Dual Student, which replaces the teacher with another student. We also define a novel concept, stable sample, following which a stabilization constraint is designed for our structure to be trainable. Further, we discuss two variants of our method, which produce even higher performance. Extensive experiments show that our method improves the classification performance significantly on several main SSL benchmarks. Specifically, it reduces the error rate of the 13-layer CNN from 16.84% to 12.39% on CIFAR-10 with 1k labels and from 34.10% to 31.56% on CIFAR-100 with 10k labels. In addition, our method also achieves a clear improvement in domain adaptation.

Very low resolution (VLR) image recognition corresponds to classifying images with resolution 16x16 or less. Though it has widespread applicability when objects are captured at a very large stand-off distance (e.g. surveillance scenario) or from wide angle mobile cameras, it has received limited attention. This research presents a novel Dual Directed Capsule Network model, termed as DirectCapsNet, for addressing VLR digit and face recognition. The proposed architecture utilizes a combination of capsule and convolutional layers for learning an effective VLR recognition model. The architecture also incorporates two novel loss functions: (i) the proposed HR-anchor loss and (ii) the proposed targeted reconstruction loss, in order to overcome the challenges of limited information content in VLR images. The proposed losses use high resolution images as auxiliary data during training to "direct" discriminative feature learning. Multiple experiments for VLR digit classification and VLR face recognition are performed along with comparisons with state-of-the-art algorithms. The proposed DirectCapsNet consistently showcases state-of-the-art results; for example, on the UCCS face database, it shows over 95% face recognition accuracy when 16x16 images are matched with 80x80 images.

Synthesizing images from a given text description involves engaging two types of information: the content, which includes information explicitly described in the text (e.g., color, composition, etc.), and the style, which is usually not well described in the text (e.g., location, quantity, size, etc.). However, in previous works, it is typically treated as a process of generating images only from the content, i.e., without considering learning meaningful style representations. In this paper, we aim to learn two variables that are disentangled in the latent space, representing content and style respectively. We achieve this by augmenting current text-to-image synthesis frameworks with a dual adversarial inference mechanism. Through extensive experiments, we show that our model learns, in an unsupervised manner, style representations corresponding to certain meaningful information present in the image that are not well described in the text. The new framework also improves the quality of synthesized images when evaluated on Oxford-102, CUB and COCO datasets.

Recent works on domain adaptation exploit adversarial training to obtain domain-invariant feature representations from the joint learning of feature extractor and domain discriminator networks. However, domain adversarial methods render suboptimal performances since they attempt to match the distributions among the domains without considering the task at hand. We propose Drop to Adapt (DTA), which leverages adversarial dropout to learn strongly discriminative features by enforcing the cluster assumption. Accordingly, we design objective functions to support robust domain adaptation. We demonstrate efficacy of the proposed method on various experiments and achieve consistent improvements in both image classification and semantic segmentation tasks. Our source code is available at https://github.com/postBG/DTA.pytorch.

In natural images, information is conveyed at different frequencies where higher frequencies are usually encoded with fine details and lower frequencies are usually encoded with global structures. Similarly, the output feature maps of a convolution layer can also be seen as a mixture of information at different frequencies. In this work, we propose to factorize the mixed feature maps by their frequencies, and design a novel Octave Convolution (OctConv) operation to store and process feature maps that vary spatially "slower" at a lower spatial resolution reducing both memory and computation cost. Unlike existing multi-scale methods, OctConv is formulated as a single, generic, plug-and-play convolutional unit that can be used as a direct replacement of (vanilla) convolutions without any adjustments in the network architecture. It is also orthogonal and complementary to methods that suggest better topologies or reduce channel-wise redundancy like group or depth-wise convolutions. We experimentally show that by simply replacing convolutions with OctConv, we can consistently boost accuracy for both image and video recognition tasks, while reducing memory and computational cost. An OctConv-equipped ResNet-152 can achieve 82.9% top-1 classification accuracy on ImageNet with merely 22.2 GFLOPs.

We introduce the novel domain-specific Drive&Act benchmark for fine-grained categorization of driver behavior. Our dataset features twelve hours and over 9.6 million frames of people engaged in distractive activities during both, manual and automated driving. We capture color, infrared, depth and 3D body pose information from six views and densely label the videos with a hierarchical annotation scheme, resulting in 83 categories. The key challenges of our dataset are: (1) recognition of fine-grained behavior inside the vehicle cabin; (2) multi-modal activity recognition, focusing on diverse data streams; and (3) a cross view recognition benchmark, where a model handles data from an unfamiliar domain, as sensor type and placement in the cabin can change between vehicles. Finally, we provide challenging benchmarks by adopting prominent methods for video- and body pose-based action recognition.

We propose to harness the potential of simulation for semantic segmentation of real-world self-driving scenes in a domain generalization fashion. The segmentation network is trained without any information about target domains and tested on the unseen target domains. To this end, we propose a new approach of domain randomization and pyramid consistency to learn a model with high generalizability. First, we propose to randomize the synthetic images with styles of real images in terms of visual appearances using auxiliary datasets, in order to effectively learn domain-invariant representations. Second, we further enforce pyramid consistency across different "stylized" images and within an image, in order to learn domain-invariant and scale-invariant features, respectively. Extensive experiments are conducted on generalization from GTA and SYNTHIA to Cityscapes, BDDS, and Mapillary; and our method achieves superior results over the state-of-the-art techniques. Remarkably, our generalization results are on par with or even better than those obtained by state-of-the-art simulation-to-real domain adaptation methods, which access the target domain data at training time.

We present a method for recovering the shared content between two visual domains as well as the content that is unique to each domain. This allows us to map from one domain to the other, in a way in which the content that is specific for the first domain is removed and the content that is specific for the second is imported from any image in the second domain. In addition, our method enables generation of images from the intersection of the two domains as well as their union, despite having no such samples during training. The method is shown analytically to contain all the sufficient and necessary constraints. It also outperforms the literature methods in an extensive set of experiments.

Deep neural networks for semantic segmentation always require a large number of samples with pixel-level labels, which becomes the major difficulty in their real-world applications. To reduce the labeling cost, unsupervised domain adaptation (UDA) approaches are proposed to transfer knowledge from labeled synthesized datasets to unlabeled real-world datasets. Recently, some semi-supervised learning methods have been applied to UDA and achieved state-of-the-art performance. One of the most popular approaches in semi-supervised learning is the entropy minimization method. However, when applying the entropy minimization to UDA for semantic segmentation, the gradient of the entropy is biased towards samples that are easy to transfer. To balance the gradient of well-classified target samples, we propose the maximum squares loss. Our maximum squares loss prevents the training process being dominated by easy-to-transfer samples in the target domain. Besides, we introduce the image-wise weighting ratio to alleviate the class imbalance in the unlabeled target domain. Both synthetic-to-real and cross-city adaptation experiments demonstrate the effectiveness of our proposed approach. The code is released at https://github. com/ZJULearning/MaxSquareLoss.

Few-shot classification consists of learning a predictive model that is able to effectively adapt to a new class, given only a few annotated samples. To solve this challenging problem, meta-learning has become a popular paradigm that advocates the ability to "learn to adapt". Recent works have shown, however, that simple learning strategies without meta-learning could be competitive. In this paper, we go a step further and show that by addressing the fundamental high-variance issue of few-shot learning classifiers, it is possible to significantly outperform current meta-learning techniques. Our approach consists of designing an ensemble of deep networks to leverage the variance of the classifiers, and introducing new strategies to encourage the networks to cooperate, while encouraging prediction diversity. Evaluation is conducted on the mini-ImageNet, tiered-ImageNet and CUB datasets, where we show that even a single network obtained by distillation yields state-of-the-art results.

Most existing methods for conditional image synthesis are only able to generate a single plausible image for any given input, or at best a fixed number of plausible images. In this paper, we focus on the problem of generating images from semantic segmentation maps and present a simple new method that can generate an arbitrary number of images with diverse appearance for the same semantic layout. Unlike most existing approaches which adopt the GAN framework, our method is based on the recently introduced Implicit Maximum Likelihood Estimation (IMLE) framework. Compared to the leading approach, our method is able to generate more diverse images while producing fewer artifacts despite using the same architecture. The learned latent space also has sensible structure despite the lack of supervision that encourages such behaviour.

This paper presents a novel method to distill knowledge from a deep pose regressor network for efficient Visual Odometry (VO). Standard distillation relies on "dark knowledge" for successful knowledge transfer. As this knowledge is not available in pose regression and the teacher prediction is not always accurate, we propose to emphasize the knowledge transfer only when we trust the teacher. We achieve this by using teacher loss as a confidence score which places variable relative importance on the teacher prediction. We inject this confidence score to the main training task via Attentive Imitation Loss (AIL) and when learning the intermediate representation of the teacher through Attentive Hint Training (AHT) approach. To the best of our knowledge, this is the first work which successfully distill the knowledge from a deep pose regression network. Our evaluation on the KITTI and Malaga dataset shows that we can keep the student prediction close to the teacher with up to 92.95% parameter reduction and 2.12x faster in computation time.

We propose to learn a 3D pose estimator by distilling knowledge from Non-Rigid Structure from Motion (NRSfM). Our method uses solely 2D landmark annotations. No 3D data, multi-view/temporal footage, or object specific prior is required. This alleviates the data bottleneck, which is one of the major concern for supervised methods. The challenge for using NRSfM as teacher is that they often make poor depth reconstruction when the 2D projections have strong ambiguity. Directly using those wrong depth as hard target would negatively impact the student. Instead, we propose a novel loss that ties depth prediction to the cost function used in NRSfM. This gives the student pose estimator freedom to reduce depth error by associating with image features. Validated on H3.6M dataset, our learned 3D pose estimation network achieves more accurate reconstruction compared to NRSfM methods. It also outperforms other weakly supervised methods, in spite of using significantly less supervision.

Video recognition models have progressed significantly over the past few years, evolving from shallow classifiers trained on hand-crafted features to deep spatiotemporal networks. However, labeled video data required to train such models has not been able to keep up with the ever increasing depth and sophistication of these networks. In this work we propose an alternative approach to learning video representations that requires no semantically labeled videos, and instead leverages the years of effort in collecting and labeling large and clean still-image datasets. We do so by using state-of-the-art models pre-trained on image datasets as "teachers" to train video models in a distillation framework. We demonstrate that our method learns truly spatiotemporal features, despite being trained only using supervision from still-image networks. Moreover, it learns good representations across different input modalities, using completely uncurated raw video data sources and with different 2D teacher models. Our method obtains strong transfer performance, outperforming standard techniques for bootstrapping video architectures with image based models by 16%. We believe that our approach opens up new approaches for learning spatiotemporal representations from unlabeled video data.

A dynamic scene has two types of elements: those that move fluidly and can be predicted from previous frames, and those which are disoccluded (exposed) and cannot be extrapolated. Prior approaches to video prediction typically learn either to warp or to hallucinate future pixels, but not both. In this paper, we describe a computational model for high-fidelity video prediction which disentangles motion-specific propagation from motion-agnostic generation. We introduce a confidence-aware warping operator which gates the output of pixel predictions from a flow predictor for non-occluded regions and from a context encoder for occluded regions. Moreover, in contrast to prior works where confidence is jointly learned with flow and appearance using a single network, we compute confidence after a warping step, and employ a separate network to inpaint exposed regions. Empirical results on both synthetic and real datasets show that our disentangling approach provides better occlusion maps and produces both sharper and more realistic predictions compared to strong baselines.