ICCV 2019 Papers

A complex visual navigation task puts an agent in different situations which call for a diverse range of visual perception abilities. For example, to "go to the nearest chair", the agent might need to identify a chair in a living room using semantics, follow along a hallway using vanishing point cues, and avoid obstacles using depth. Therefore, utilizing the appropriate visual perception abilities based on a situational understanding of the visual environment can empower these navigation models in unseen visual environments. We propose to train an agent to fuse a large set of visual representations that correspond to diverse visual perception abilities. To fully utilize each representation, we develop an action-level representation fusion scheme, which predicts an action candidate from each representation and adaptively consolidate these action candidates into the final action. Furthermore, we employ a data-driven inter-task affinity regularization to reduce redundancies and improve generalization. Our approach leads to a significantly improved performance in novel environments over ImageNet-pretrained baseline and other fusion methods.

Multi-person pose estimation is a challenging problem. Existing methods are mostly two-stage based-one stage for proposal generation and the other for allocating poses to corresponding persons. However, such two-stage methods generally suffer low efficiency. In this work, we present the first single-stage model, Single-stage multi-person Pose Machine (SPM), to simplify the pipeline and lift the efficiency for multi-person pose estimation. To achieve this, we propose a novel Structured Pose Representation (SPR) that unifies person instance and body joint position representations. Based on SPR, we develop the SPM model that can directly predict structured poses for multiple persons in a single stage, and thus offer a more compact pipeline and attractive efficiency advantage over two-stage methods. In particular, SPR introduces the root joints to indicate different person instances and human body joint positions are encoded into their displacements w.r.t. the roots. To better predict long-range displacements for some joints, SPR is further extended to hierarchical representations. Based on SPR, SPM can efficiently perform multi-person poses estimation by simultaneously predicting root joints (location of instances) and body joint displacements via CNNs. Moreover, to demonstrate the generality of SPM, we also apply it to multi-person 3D pose estimation. Comprehensive experiments on benchmarks MPII, extended PASCAL-Person-Part, MSCOCO and CMU Panoptic clearly demonstrate the state-of-the-art efficiency of SPM for multi-person 2D/3D pose estimation, together with outstanding accuracy.

We present the first single-network approach for 2D whole-body pose estimation, which entails simultaneous localization of body, face, hands, and feet keypoints. Due to the bottom-up formulation, our method maintains constant real-time performance regardless of the number of people in the image. The network is trained in a single stage using multi-task learning, through an improved architecture which can handle scale differences between body/foot and face/hand keypoints. Our approach considerably improves upon OpenPose [??], the only work so far capable of whole-body pose estimation, both in terms of speed and global accuracy. Unlike OpenPose, our method does not need to run an additional network for each hand and face candidate, making it substantially faster for multi-person scenarios. This work directly results in a reduction of computational complexity for applications that require 2D whole-body information (e.g., VR/AR, re-targeting). In addition, it yields higher accuracy, especially for occluded, blurry, and low resolution faces and hands. For code, trained models, and validation benchmarks, visit our project page: https://github.com/CMU-Perceptual-Computing-Lab/openpose_train.

We introduce SinGAN, an unconditional generative model that can be learned from a single natural image. Our model is trained to capture the internal distribution of patches within the image, and is then able to generate high quality, diverse samples that carry the same visual content as the image. SinGAN contains a pyramid of fully convolutional GANs, each responsible for learning the patch distribution at a different scale of the image. This allows generating new samples of arbitrary size and aspect ratio, that have significant variability, yet maintain both the global structure and the fine textures of the training image. In contrast to previous single image GAN schemes, our approach is not limited to texture images, and is not conditional (i.e. it generates samples from noise). User studies confirm that the generated samples are commonly confused to be real images. We illustrate the utility of SinGAN in a wide range of image manipulation tasks.

We propose to jointly learn multi-view geometry and warping between views of the same object instances for robust cross-view object detection. What makes multi-view object instance detection difficult are strong changes in viewpoint, lighting conditions, high similarity of neighbouring objects, and strong variability in scale. By turning object detection and instance re-identification in different views into a joint learning task, we are able to incorporate both image appearance and geometric soft constraints into a single, multi-view detection process that is learnable end-to-end. We validate our method on a new, large data set of street-level panoramas of urban objects and show superior performance compared to various baselines. Our contribution is threefold: a large-scale, publicly available data set for multi-view instance detection and re-identification; an annotation tool custom-tailored for multi-view instance detection; and a novel, holistic multi-view instance detection and re-identification method that jointly models geometry and appearance across views.

Knowledge distillation is a widely applicable technique for training a student neural network under the guidance of a trained teacher network. For example, in neural network compression, a high-capacity teacher is distilled to train a compact student; in privileged learning, a teacher trained with privileged data is distilled to train a student without access to that data. The distillation loss determines how a teacher's knowledge is captured and transferred to the student. In this paper, we propose a new form of knowledge distillation loss that is inspired by the observation that semantically similar inputs tend to elicit similar activation patterns in a trained network. Similarity-preserving knowledge distillation guides the training of a student network such that input pairs that produce similar (dissimilar) activations in the teacher network produce similar (dissimilar) activations in the student network. In contrast to previous distillation methods, the student is not required to mimic the representation space of the teacher, but rather to preserve the pairwise similarities in its own representation space. Experiments on three public datasets demonstrate the potential of our approach.

For unsupervised domain adaptation problems, the strategy of aligning the two domains in latent feature space through adversarial learning has achieved much progress in image classification, but usually fails in semantic segmentation tasks in which the latent representations are overcomplex. In this work, we equip the adversarial network with a "significance-aware information bottleneck (SIB)", to address the above problem. The new network structure, called SIBAN, enables a significance-aware feature purification before the adversarial adaptation, which eases the feature alignment and stabilizes the adversarial training course. In two domain adaptation tasks, i.e., GTA5 -> Cityscapes and SYNTHIA -> Cityscapes, we validate that the proposed method can yield leading results compared with other feature-space alternatives. Moreover, SIBAN can even match the state-of-the-art output-space methods in segmentation accuracy, while the latter are often considered to be better choices for domain adaptive segmentation task.

Siamese networks are non-linear deep models that have found their ways into a broad set of problems in learning theory, thanks to their embedding capabilities. In this paper, we study Siamese networks from a new perspective and question the validity of their training procedure. We show that in the majority of cases, the objective of a Siamese network is endowed with an invariance property. Neglecting the invariance property leads to a hindrance in training the Siamese networks. To alleviate this issue, we propose two Riemannian structures and generalize a well-established accelerated stochastic gradient descent method to take into account the proposed Riemannian structures. Our empirical evaluations suggest that by making use of the Riemannian geometry, we achieve state-of-the-art results against several algorithms for the challenging problem of fine-grained image classification.

Deep learning with 3D data has progressed significantly since the introduction of convolutional neural networks that can handle point order ambiguity in point cloud data. While being able to achieve good accuracies in various scene understanding tasks, previous methods often have low training speed and complex network architecture. In this paper, we address these problems by proposing an efficient end-to-end permutation invariant convolution for point cloud deep learning. Our simple yet effective convolution operator named ShellConv uses statistics from concentric spherical shells to define representative features and resolve the point order ambiguity, allowing traditional convolution to perform on such features. Based on ShellConv we further build an efficient neural network named ShellNet to directly consume the point clouds with larger receptive fields while maintaining less layers. We demonstrate the efficacy of ShellNet by producing state-of-the-art results on object classification, object part segmentation, and semantic scene segmentation while keeping the network very fast to train.

Recent developments in gradient-based attention modeling have seen attention maps emerge as a powerful tool for interpreting convolutional neural networks. Despite good localization for an individual class of interest, these techniques produce attention maps with substantially overlapping responses among different classes, leading to the problem of visual confusion and the need for discriminative attention. In this paper, we address this problem by means of a new framework that makes class-discriminative attention a principled part of the learning process. Our key innovations include new learning objectives for attention separability and cross-layer consistency, which result in improved attention discriminability and reduced visual confusion. Extensive experiments on image classification benchmarks show the effectiveness of our approach in terms of improved classification accuracy, including CIFAR-100 (+3.33%), Caltech-256 (+1.64%), ImageNet (+0.92%), CUB-200-2011 (+4.8%) and PASCAL VOC2012 (+5.73%).

In this work we explore how fine-grained differences between the shapes of common objects are expressed in language, grounded on 2D and/or 3D object representations. We first build a large scale, carefully controlled dataset of human utterances each of which refers to a 2D rendering of a 3D CAD model so as to distinguish it from a set of shape-wise similar alternatives. Using this dataset, we develop neural language understanding (listening) and production (speaking) models that vary in their grounding (pure 3D forms via point-clouds vs. rendered 2D images), the degree of pragmatic reasoning captured (e.g. speakers that reason about a listener or not), and the neural architecture (e.g. with or without attention). We find models that perform well with both synthetic and human partners, and with held out utterances and objects. We also find that these models are capable of zero-shot transfer learning to novel object classes (e.g. transfer from training on chairs to testing on lamps), as well as to real-world images drawn from furniture catalogs. Lesion studies indicate that the neural listeners depend heavily on part-related words and associate these words correctly with visual parts of objects (without any explicit supervision on such parts), and that transfer to novel classes is most successful when known part-related words are available. This work illustrates a practical approach to language grounding, and provides a novel case study in the relationship between object shape and linguistic structure when it comes to object differentiation.

Instance segmentation aims to detect and segment individual objects in a scene. Most existing methods rely on precise mask annotations of every category. However, it is difficult and costly to segment objects in novel categories because a large number of mask annotations is required. We introduce ShapeMask, which learns the intermediate concept of object shape to address the problem of generalization in instance segmentation to novel categories. ShapeMask starts with a bounding box detection and gradually refines it by first estimating the shape of the detected object through a collection of shape priors. Next, ShapeMask refines the coarse shape into an instance level mask by learning instance embeddings. The shape priors provide a strong cue for object-like prediction, and the instance embeddings model the instance specific appearance information. ShapeMask significantly outperforms the state-of-the-art by 6.4 and 3.8 AP when learning across categories, and obtains competitive performance in the fully supervised setting. It is also robust to inaccurate detections, decreased model capacity, and small training data. Moreover, it runs efficiently with 150ms inference time on a GPU and trains within 11 hours on TPUs. With a larger backbone model, ShapeMask increases the gap with state-of-the-art to 9.4 and 6.2 AP across categories. Code will be publicly available at: https://sites.google.com/view/shapemask/home.

We propose a scalable neural network framework to reconstruct the 3D mesh of a human body from multi-view images, in the subspace of the SMPL model. Use of multi-view images can significantly reduce the projection ambiguity of the problem, increasing the reconstruction accuracy of the 3D human body under clothing. Our experiments show that this method benefits from the synthetic dataset generated from our pipeline since it has good flexibility of variable control and can provide ground-truth for validation. Our method outperforms existing methods on real-world images, especially on shape estimations.

We investigate the problem of reconstructing shapes from noisy and incomplete projections in the presence of viewpoint uncertainities. The problem is cast as an optimization over the shape given measurements obtained by a projection operator and a prior. We present differentiable projection operators for a number of reconstruction problems which when combined with the deep image prior or shape prior allows efficient inference through gradient descent. We apply our method on a variety of reconstruction problems, such as tomographic reconstruction from a few samples, visual hull reconstruction incorporating view uncertainties, and 3D shape reconstruction from noisy depth maps. Experimental results show that our approach is effective for such shape reconstruction problems, without requiring any task-specific training.

We propose a novel deep learning method for shadow removal. Inspired by physical models of shadow formation, we use a linear illumination transformation to model the shadow effects in the image that allows the shadow image to be expressed as a combination of the shadow-free image, the shadow parameters, and a matte layer. We use two deep networks, namely SP-Net and M-Net, to predict the shadow parameters and the shadow matte respectively. This system allows us to remove the shadow effects on the images. We train and test our framework on the most challenging shadow removal dataset (ISTD). Compared to the state-of-the-art method, our model achieves a 40% error reduction in terms of root mean square error (RMSE) for the shadow area, reducing RMSE from 13.3 to 7.9. Moreover, we create an augmented ISTD dataset based on an image decomposition system by modifying the shadow parameters to generate new synthetic shadow images. Training our model on this new augmented ISTD dataset further lowers the RMSE on the shadow area to 7.4.

Diverse and accurate vision+language modeling is an important goal to retain creative freedom and maintain user engagement. However, adequately capturing the intricacies of diversity in language models is challenging. Recent works commonly resort to latent variable models augmented with more or less supervision from object detectors or part-of-speech tags. In common to all those methods is the fact that the latent variable either only initializes the sentence generation process or is identical across the steps of generation. Both methods offer no fine-grained control. To address this concern, we propose Seq-CVAE which learns a latent space for every word. We encourage this temporal latent space to capture the 'intention' about how to complete the sentence by mimicking a representation which summarizes the future. We illustrate the efficacy of the proposed approach on the challenging MSCOCO dataset, significantly improving diversity metrics compared to baselines while performing on par w.r.t. sentence quality.

We propose a self-supervised learning framework for visual odometry (VO) that incorporates correlation of consecutive frames and takes advantage of adversarial learning. Previous methods tackle self-supervised VO as a local structure from motion (SfM) problem that recovers depth from single image and relative poses from image pairs by minimizing photometric loss between warped and captured images. As single-view depth estimation is an ill-posed problem, and photometric loss is incapable of discriminating distortion artifacts of warped images, the estimated depth is vague and pose is inaccurate. In contrast to previous methods, our framework learns a compact representation of frame-to-frame correlation, which is updated by incorporating sequential information. The updated representation is used for depth estimation. Besides, we tackle VO as a self-supervised image generation task and take advantage of Generative Adversarial Networks (GAN). The generator learns to estimate depth and pose to generate a warped target image. The discriminator evaluates the quality of generated image with high-level structural perception that overcomes the problem of pixel-wise loss in previous methods. Experiments on KITTI and Cityscapes datasets show that our method obtains more accurate depth with details preserved and predicted pose outperforms state-of-the-art self-supervised methods significantly.

Video objection detection (VID) has been a rising research direction in recent years. A central issue of VID is the appearance degradation of video frames caused by fast motion. This problem is essentially ill-posed for a single frame. Therefore, aggregating features from other frames becomes a natural choice. Existing methods rely heavily on optical flow or recurrent neural networks for feature aggregation. However, these methods emphasize more on the temporally nearby frames. In this work, we argue that aggregating features in the full-sequence level will lead to more discriminative and robust features for video object detection. To achieve this goal, we devise a novel Sequence Level Semantics Aggregation (SELSA) module. We further demonstrate the close relationship between the proposed method and the classic spectral clustering method, providing a novel view for understanding the VID problem. We test the proposed method on the ImageNet VID and the EPIC KITCHENS dataset and achieve new state-of-the-art results. Our method does not need complicated postprocessing methods such as Seq-NMS or Tubelet rescoring, which keeps the pipeline simple and clean.

Generating semantic layout from scene graph is a crucial intermediate task connecting text to image. We present a conceptually simple, flexible and general framework using sequence to sequence (seq-to-seq) learning for this task. The framework, called Seq-SG2SL, derives sequence proxies for the two modality and a Transformer-based seq-to-seq model learns to transduce one into the other. A scene graph is decomposed into a sequence of semantic fragments (SF), one for each relationship. A semantic layout is represented as the consequence from a series of brick-action code segments (BACS), dictating the position and scale of each object bounding box in the layout. Viewing the two building blocks, SF and BACS, as corresponding terms in two different vocabularies, a seq-to-seq model is fittingly used to translate. A new metric, semantic layout evaluation understudy (SLEU), is devised to evaluate the task of semantic layout prediction inspired by BLEU. SLEU defines relationships within a layout as unigrams and looks at the spatial distribution for n-grams. Unlike the binary precision of BLEU, SLEU allows for some tolerances spatially through thresholding the Jaccard Index and is consequently more adapted to the task. Experimental results on the challenging Visual Genome dataset show improvement over a non-sequential approach based on graph convolution.