ICCV 2019 Papers

Despite the great progress made by deep CNNs in image semantic segmentation, they typically require a large number of densely-annotated images for training and are difficult to generalize to unseen object categories. Few-shot segmentation has thus been developed to learn to perform segmentation from only a few annotated examples. In this paper, we tackle the challenging few-shot segmentation problem from a metric learning perspective and present PANet, a novel prototype alignment network to better utilize the information of the support set. Our PANet learns class-specific prototype representations from a few support images within an embedding space and then performs segmentation over the query images through matching each pixel to the learned prototypes. With non-parametric metric learning, PANet offers high-quality prototypes that are representative for each semantic class and meanwhile discriminative for different classes. Moreover, PANet introduces a prototype alignment regularization between support and query. With this, PANet fully exploits knowledge from the support and provides better generalization on few-shot segmentation. Significantly, our model achieves the mIoU score of 48.1% and 55.7% on PASCAL-5i for 1-shot and 5-shot settings respectively, surpassing the state-of-the-art method by 1.8% and 8.6%.

In comparison with person re-identification (ReID), which has been widely studied in the research community, vehicle ReID has received less attention. Vehicle ReID is challenging due to 1) high intra-class variability (caused by the dependency of shape and appearance on viewpoint), and 2) small inter-class variability (caused by the similarity in shape and appearance between vehicles produced by different manufacturers). To address these challenges, we propose a Pose-Aware Multi-Task Re-Identification (PAMTRI) framework. This approach includes two innovations compared with previous methods. First, it overcomes viewpoint-dependency by explicitly reasoning about vehicle pose and shape via keypoints, heatmaps and segments from pose estimation. Second, it jointly classifies semantic vehicle attributes (colors and types) while performing ReID, through multi-task learning with the embedded pose representations. Since manually labeling images with detailed pose and attribute information is prohibitive, we create a large-scale highly randomized synthetic dataset with automatically annotated vehicle attributes for training. Extensive experiments validate the effectiveness of each proposed component, showing that PAMTRI achieves significant improvement over state-of-the-art on two mainstream vehicle ReID benchmarks: VeRi and CityFlow-ReID.

Learning-based methods are demonstrating their strong competitiveness in estimating depth for multi-view stereo reconstruction in recent years. Among them the approaches that generate cost volumes based on the plane-sweeping algorithm and then use them for feature matching have shown to be very prominent recently. The plane-sweep volumes are essentially anisotropic in depth and spatial directions, but they are often approximated by isotropic cost volumes in those methods, which could be detrimental. In this paper, we propose a new end-to-end deep learning network of P-MVSNet for multi-view stereo based on isotropic and anisotropic 3D convolutions. Our P-MVSNet consists of two core modules: a patch-wise aggregation module learns to aggregate the pixel-wise correspondence information of extracted features to generate a matching confidence volume, from which a hybrid 3D U-Net then infers a depth probability distribution and predicts the depth maps. We perform extensive experiments on the DTU and Tanks & Temples benchmark datasets, and the results show that the proposed P-MVSNet achieves the state-of-the-art performance over many existing methods on multi-view stereo.

Lifelong learning with deep neural networks is well-known to suffer from catastrophic forgetting: the performance on previous tasks drastically degrades when learning a new task. To alleviate this effect, we propose to leverage a large stream of unlabeled data easily obtainable in the wild. In particular, we design a novel class-incremental learning scheme with (a) a new distillation loss, termed global distillation, (b) a learning strategy to avoid overfitting to the most recent task, and (c) a confidence-based sampling method to effectively leverage unlabeled external data. Our experimental results on various datasets, including CIFAR and ImageNet, demonstrate the superiority of the proposed methods over prior methods, particularly when a stream of unlabeled data is accessible: our method shows up to 15.8% higher accuracy and 46.5% less forgetting compared to the state-of-the-art method. The code is available at https://github.com/kibok90/iccv2019-inc.

We address semantic segmentation on omnidirectional images, to leverage a holistic understanding of the surrounding scene for applications like autonomous driving systems. For the spherical domain, several methods recently adopt an icosahedron mesh, but systems are typically rotation invariant or require significant memory and parameters, thus enabling execution only at very low resolutions. In our work, we propose an orientation-aware CNN framework for the icosahedron mesh. Our representation allows for fast network operations, as our design simplifies to standard network operations of classical CNNs, but under consideration of north-aligned kernel convolutions for features on the sphere. We implement our representation and demonstrate its memory efficiency up-to a level-8 resolution mesh (equivalent to 640 x 1024 equirectangular images). Finally, since our kernels operate on the tangent of the sphere, standard feature weights, pretrained on perspective data, can be directly transferred with only small need for weight refinement. In our evaluation our orientation-aware CNN becomes a new state of the art for the recent 2D3DS dataset, and our Omni-SYNTHIA version of SYNTHIA. Rotation invariant classification and segmentation tasks are additionally presented for comparison to prior art.

Supervised dimensionality reduction for sequence data projects the observations in sequences onto a low-dimensional subspace to better separate different sequence classes. It is typically more challenging than conventional dimensionality reduction for static data, because measuring the separability of sequences involves non-linear procedures to manipulate the temporal structures. This paper presents a linear method, namely Order-preserving Wasserstein Discriminant Analysis (OWDA), which learns the projection by maximizing the inter-class distance and minimizing the intra-class scatter. For each class, OWDA extracts the order-preserving Wasserstein barycenter and constructs the intra-class scatter as the dispersion of the training sequences around the barycenter. The inter-class distance is measured as the order-preserving Wasserstein distance between the corresponding barycenters. OWDA is able to concentrate on the distinctive differences among classes by lifting the geometric relations with temporal constraints. Experiments show that OWDA achieves competitive results on three 3D action recognition datasets.

In this paper, we study the problem of sequentially building houses in the game of Minecraft, and demonstrate that learning the ordering can make for more effective autoregressive models. Given a partially built house made by a human player, our system tries to place additional blocks in a human-like manner to complete the house. We introduce a new dataset, HouseCraft, for this new task. HouseCraft contains the sequential order in which 2,500 Minecraft houses were built from scratch by humans. The human action sequences enable us to learn an order-aware generative model called Voxel-CNN. In contrast to many generative models where the sequential generation ordering either does not matter (e.g. holistic generation with GANs), or is manually/arbitrarily set by simple rules (e.g. raster-scan order), our focus is on an ordered generation that imitates humans. To evaluate if a generative model can accurately predict human-like actions, we propose several novel quantitative metrics. We demonstrate that our Voxel-CNN model is simple and effective at this creative task, and can serve as a strong baseline for future research in this direction. The HouseCraft dataset and code with baseline models will be made publicly available.

Current CNN-based solutions to salient object detection (SOD) mainly rely on the optimization of cross-entropy loss (CELoss). Then the quality of detected saliency maps is often evaluated in terms of F-measure. In this paper, we investigate an interesting issue: can we consistently use the F-measure formulation in both training and evaluation for SOD? By reformulating the standard F-measure we propose the relaxed F-measure which is differentiable w.r.t the posterior and can be easily appended to the back of CNNs as the loss function. Compared to the conventional cross-entropy loss of which the gradients decrease dramatically in the saturated area, our loss function, named FLoss, holds considerable gradients even when the activation approaches the target. Consequently, the FLoss can continuously force the network to produce polarized activations. Comprehensive benchmarks on several popular datasets show that FLoss outperforms the state-of-the-art with a considerable margin. More specifically, due to the polarized predictions, our method is able to obtain high-quality saliency maps without carefully tuning the optimal threshold, showing significant advantages in real-world applications.

A human pose is naturally represented as a graph where the joints are the nodes and the bones are the edges. So it is natural to apply Graph Convolutional Network (GCN) to estimate 3D poses from 2D poses. In this work, we propose a generic formulation where both GCN and Fully Connected Network (FCN) are its special cases. From this formulation, we discover that GCN has limited representation power when used for estimating 3D poses. We overcome the limitation by introducing Locally Connected Network (LCN) which is naturally implemented by this generic formulation. It notably improves the representation capability over GCN. In addition, since every joint is only connected to a few joints in its neighborhood, it has strong generalization power. The experiments on public datasets show it: (1) outperforms the state-of-the-arts; (2) is less data hungry than alternative models; (3) generalizes well to unseen actions and datasets.

This paper proposes a learning-based framework for reconstructing 3D shapes from functional operators, compactly encoded as small-sized matrices. To this end we introduce a novel neural architecture, called OperatorNet, which takes as input a set of linear operators representing a shape and produces its 3D embedding. We demonstrate that this approach significantly outperforms previous purely geometric methods for the same problem. Furthermore, we introduce a novel functional operator, which encodes the extrinsic or pose-dependent shape information, and thus complements purely intrinsic pose-oblivious operators, such as the classical Laplacian. Coupled with this novel operator, our reconstruction network achieves very high reconstruction accuracy, even in the presence of incomplete information about a shape, given a soft or functional map expressed in a reduced basis. Finally, we demonstrate that the multiplicative functional algebra enjoyed by these operators can be used to synthesize entirely new unseen shapes, in the context of shape interpolation and shape analogy applications.

This work addresses the problem of 3D kinematic structure learning of arbitrary articulated rigid bodies from RGB-D data sequences. Typically, this problem is addressed by offline methods that process a batch of frames, assuming that complete point trajectories are available. However, this approach is not feasible when considering scenarios that require continuity and fluidity, for instance, human-robot interaction. In contrast, we propose to tackle this problem in an online unsupervised fashion, by recursively maintaining the metric distance of the scene's 3D structure, while achieving real-time performance. The influence of noise is mitigated by building a similarity measure based on a linear embedding representation and incorporating this representation into the original metric distance. The kinematic structure is then estimated based on a combination of implicit motion and spatial properties. The proposed approach achieves competitive performance both quantitatively and qualitatively in terms of estimation accuracy, even compared to offline methods.

High-quality computer vision models typically address the problem of understanding the general distribution of real-world images. However, most cameras observe only a very small fraction of this distribution. This offers the possibility of achieving more efficient inference by specializing compact, low-cost models to the specific distribution of frames observed by a single camera. In this paper, we employ the technique of model distillation (supervising a low-cost student model using the output of a high-cost teacher) to specialize accurate, low-cost semantic segmentation models to a target video stream. Rather than learn a specialized student model on offline data from the video stream, we train the student in an online fashion on the live video, intermittently running the teacher to provide a target for learning. Online model distillation yields semantic segmentation models that closely approximate their Mask R-CNN teacher with 7 to 17xlower inference runtime cost (11 to 26xin FLOPs), even when the target video's distribution is non-stationary. Our method requires no offline pretraining on the target video stream, achieves higher accuracy and lower cost than solutions based on flow or video object segmentation, and can exhibit better temporal stability than the original teacher. We also provide a new video dataset for evaluating the efficiency of inference over long running video streams.

Data augmentation is critical to the success of modern deep learning techniques. In this paper, we propose Online Hyper-parameter Learning for Auto-Augmentation (OHL-Auto-Aug), an economical solution that learns the augmentation policy distribution along with network training. Unlike previous methods on auto-augmentation that search augmentation strategies in an offline manner, our method formulates the augmentation policy as a parameterized probability distribution, thus allowing its parameters to be optimized jointly with network parameters. Our proposed OHL-Auto-Aug eliminates the need of re-training and dramatically reduces the cost of the overall search process, while establishes significantly accuracy improvements over baseline models. On both CIFAR-10 and ImageNet, our method achieves remarkable on search accuracy, 60x faster on CIFAR-10 and 24x faster on ImageNet, while maintaining competitive accuracies.

We propose the onion-peel networks for video completion. Given a set of reference images and a target image with holes, our network fills the hole by referring the contents in the reference images. Our onion-peel network progressively fills the hole from the hole boundary enabling it to exploit richer contextual information for the missing regions every step. Given a sufficient number of recurrences, even a large hole can be inpainted successfully. To attend to the missing information visible in the reference images, we propose an asymmetric attention block that computes similarities between the hole boundary pixels in the target and the non-hole pixels in the references in a non-local manner. With our attention block, our network can have an unlimited spatial-temporal window size and fill the holes with globally coherent contents. In addition, our framework is applicable to the image completion guided by the reference images without any modification, which is difficult to do with the previous methods. We validate that our method produces visually pleasing image and video inpainting results in realistic test cases.

Neural architecture search (NAS) aims to automate the search procedure of architecture instead of manual design. Even if recent NAS approaches finish the search within days, lengthy training is still required for a specific architecture candidate to get the parameters for its accurate evaluation. Recently one-shot NAS methods are proposed to largely squeeze the tedious training process by sharing parameters across candidates. In this way, the parameters for each candidate can be directly extracted from the shared parameters instead of training them from scratch. However, they have no sense of which candidate will perform better until evaluation so that the candidates to evaluate are randomly sampled and the top-1 candidate is considered the best. In this paper, we propose a Self-Evaluated Template Network (SETN) to improve the quality of the architecture candidates for evaluation so that it is more likely to cover competitive candidates. SETN consists of two components: (1) an evaluator, which learns to indicate the probability of each individual architecture being likely to have a lower validation loss. The candidates for evaluation can thus be selectively sampled according to this evaluator. (2) a template network, which shares parameters among all candidates to amortize the training cost of generated candidates. In experiments, the architecture found by SETN achieves the state-of-the-art performance on CIFAR and ImageNet benchmarks within comparable computation costs.

Modern deep neural networks are often vulnerable to adversarial samples. Based on the first optimization-based attacking method, many following methods are proposed to improve the attacking performance and speed. Recently, generation-based methods have received much attention since they directly use feed-forward networks to generate the adversarial samples, which avoid the time-consuming iterative attacking procedure in optimization-based and gradient-based methods. However, current generation-based methods are only able to attack one specific target (category) within one model, thus making them not applicable to real classification systems that often have hundreds/thousands of categories. In this paper, we propose the first Multi-target Adversarial Network (MAN), which can generate multi-target adversarial samples with a single model. By incorporating the specified category information into the intermediate features, it can attack any category of the target classification model during runtime. Experiments show that the proposed MAN can produce stronger attack results and also have better transferability than previous state-of-the-art methods in both multi-target attack task and single-target attack task. We further use the adversarial samples generated by our MAN to improve the robustness of the classification model. It can also achieve better classification accuracy than other methods when attacked by various methods.

Currently, most deep learning based disparity estimation methods have the problem of over-smoothing at boundaries, which is unfavorable for some applications such as point cloud segmentation, mapping, etc. To address this problem, we first analyze the potential causes and observe that the estimated disparity at edge boundary pixels usually follows multimodal distributions, causing over-smoothing estimation. Based on this observation, we propose a single-modal weighted average operation on the probability distribution during inference, which can alleviate the problem effectively. To integrate the constraint of this inference method into training stage, we further analyze the characteristics of different loss functions and found that using cross entropy with gaussian distribution consistently further improves the performance. For quantitative evaluation, we propose a novel metric that measures the disparity error in the local structure of edge boundaries. Experiments on various datasets using various networks show our method's effectiveness and general applicability. Code will be available at https://github.com/chenchr/otosp.

Adversarial representation learning is a promising paradigm for obtaining data representations that are invariant to certain sensitive attributes while retaining the information necessary for predicting target attributes. Existing approaches solve this problem through iterative adversarial minimax optimization and lack theoretical guarantees. In this paper, we first study the "linear" form of this problem i.e., the setting where all the players are linear functions. We show that the resulting optimization problem is both non-convex and non-differentiable. We obtain an exact closed-form expression for its global optima through spectral learning and provide performance guarantees in terms of analytical bounds on the achievable utility and invariance. We then extend this solution and analysis to non-linear functions through kernel representation. Numerical experiments on UCI, Extended Yale B and CIFAR-100 datasets indicate that, (a) practically, our solution is ideal for "imparting" provable invariance to any biased pre-trained data representation, and (b) the global optima of the "kernel" form can provide a comparable trade-off between utility and invariance in comparison to iterative minimax optimization of existing deep neural network based approaches, but with provable guarantees.

In this paper, we present a thorough evaluation of the efficacy of knowledge distillation and its dependence on student and teacher architectures. Starting with the observation that more accurate teachers often don't make good teachers, we attempt to tease apart the factors that affect knowledge distillation performance. We find crucially that larger models do not often make better teachers. We show that this is a consequence of mismatched capacity, and that small students are unable to mimic large teachers. We find typical ways of circumventing this (such as performing a sequence of knowledge distillation steps) to be ineffective. Finally, we show that this effect can be mitigated by stopping the teacher's training early. Our results generalize across datasets and models.