CVPR 2020 Papers

We present MSeg, a composite dataset that unifies se- mantic segmentation datasets from different domains. A naive merge of the constituent datasets yields poor performance due to inconsistent taxonomies and annotation practices. We reconcile the taxonomies and bring the pixel-level annotations into alignment by relabeling more than 220,000 object masks in more than 80,000 images. The resulting composite dataset enables training a single semantic segmentation model that functions effectively across domains and generalizes to datasets that were not seen during training. We adopt zero-shot cross-dataset transfer as a benchmark to systematically evaluate a model's robustness and show that MSeg training yields substantially more robust models in comparison to training on individual datasets or naive mixing of datasets without the presented contributions. A model trained on MSeg ranks first on the WildDash leaderboard for robust semantic segmentation, with no exposure to WildDash data during training.

While Generative Adversarial Networks (GANs) have seen huge successes in image synthesis tasks, they are notoriously difficult to adapt to different datasets, in part due to instability during training and sensitivity to hyperparameters. One commonly accepted reason for this instability is that gradients passing from the discriminator to the generator become uninformative when there isn't enough overlap in the supports of the real and fake distributions. In this work, we propose the Multi-Scale Gradient Generative Adversarial Network (MSG-GAN), a simple but effective technique for addressing this by allowing the flow of gradients from the discriminator to the generator at multiple scales. This technique provides a stable approach for high resolution image synthesis, and serves as an alternative to the commonly used progressive growing technique. We show that MSG-GAN converges stably on a variety of image datasets of different sizes, resolutions and domains, as well as different types of loss functions and architectures, all with the same set of fixed hyperparameters. When compared to state-of-the-art GANs, our approach matches or exceeds the performance in most of the cases we tried.

In late fusion, each modality is processed in a separate unimodal Convolutional Neural Network (CNN) stream and the scores of each modality are fused at the end. Due to its simplicity, late fusion is still the predominant approach in many state-of-the-art multimodal applications. In this paper, we present a simple neural network module for leveraging the knowledge from multiple modalities in convolutional neural networks. The proposed unit, named Multimodal Transfer Module (MMTM), can be added at different levels of the feature hierarchy, enabling slow modality fusion. Using squeeze and excitation operations, MMTM utilizes the knowledge of multiple modalities to recalibrate the channel-wise features in each CNN stream. Unlike other intermediate fusion methods, the proposed module could be used for feature modality fusion in convolution layers with different spatial dimensions. Another advantage of the proposed method is that it could be added among unimodal branches with minimum changes in the their network architectures, allowing each branch to be initialized with existing pretrained weights. Experimental results show that our framework improves the recognition accuracy of well-known multimodal networks. We demonstrate state-of-the-art or competitive performance on four datasets that span the task domains of dynamic hand gesture recognition, speech enhancement, and action recognition with RGB and body joints.

In this paper, we address the 3D object detection task by capturing multi-level contextual information with the self-attention mechanism and multi-scale feature fusion. Most existing 3D object detection methods recognize objects individually, without giving any consideration on contextual information between these objects. Comparatively, we propose Multi-Level Context VoteNet (MLCVNet) to recognize 3D objects correlatively, building on the state-of-the-art VoteNet. We introduce three context modules into the voting and classifying stages of VoteNet to encode contextual information at different levels. Specifically, a Patch-to-Patch Context (PPC) module is employed to capture contextual information between the point patches, before voting for their corresponding object centroid points. Subsequently, an Object-to-Object Context (OOC) module is incorporated before the proposal and classification stage, to capture the contextual information between object candidates. Finally, a Global Scene Context (GSC) module is designed to learn the global scene context. We demonstrate these by capturing contextual information at patch, object and scene levels. Our method is an effective way to promote detection accuracy, achieving new state-of-the-art detection performance on challenging 3D object detection datasets, i.e., SUN RGBD and ScanNet. We also release our code at https://github.com/NUAAXQ/MLCVNet.

In this paper, we explore synthesizing person images with multiple conditions for various backgrounds. To this end, we propose a framework named "MISC" for conditional image generation and image compositing. For conditional image generation, we improve the existing condition injection mechanisms by leveraging the inter-condition correlations. For the image compositing, we theoretically prove the weaknesses of the cutting-edge methods, and make it more robust by removing the spatially-invariance constraint, and enabling the bounding mechanism and the spatial adaptability. We show the effectiveness of our method on the Video Instance-level Parsing dataset, and demonstrate the robustness through controllability tests.

We present a convex mixed-integer programming formulation for non-rigid shape matching. To this end, we propose a novel shape deformation model based on an efficient low-dimensional discrete model, so that finding a globally optimal solution is tractable in (most) practical cases. Our approach combines several favourable properties, namely it is independent of the initialisation, it is much more efficient to solve to global optimality compared to analogous quadratic assignment problem formulations, and it is highly flexible in terms of the variants of matching problems it can handle. Experimentally we demonstrate that our approach outperforms existing methods for sparse shape matching, that it can be used for initialising dense shape matching methods, and we showcase its flexibility on several examples.

Body orientation estimation provides crucial visual cues in many applications, including robotics and autonomous driving. It is particularly desirable when 3-D pose estimation is difficult to infer due to poor image resolution, occlusion or indistinguishable body parts. We present COCO-MEBOW (Monocular Estimation of Body Orientation in the Wild), a new large-scale dataset for orientation estimation from a single in-the-wild image. The body-orientation labels for around 130K human bodies within 55K images from the COCO dataset have been collected using an efficient and high-precision annotation pipeline. We also validated the benefits of the dataset. First, we show that our dataset can substantially improve the performance and the robustness of a human body orientation estimation model, the development of which was previously limited by the scale and diversity of the available training data. Additionally, we present a novel triple-source solution for 3-D human pose estimation, where 3-D pose labels, 2-D pose labels, and our body-orientation labels are all used in joint training. Our model significantly outperforms state-of-the-art dual-source solutions for monocular 3-D human pose estimation, where training only uses 3-D pose labels and 2-D pose labels. This substantiates an important advantage of MEBOW for 3-D human pose estimation, which is particularly appealing because the per-instance labeling cost for body orientations is far less than that for 3-D poses. The work demonstrates high potential of MEBOW in addressing real-world challenges involving understanding human behaviors. Further information of this work is available at https://chenyanwu.github.io/MEBOW/.

Nowadays, driven by the increasing concern on diet and health, food computing has attracted enormous attention from both industry and research community. One of the most popular research topics in this domain is Food Retrieval, due to its profound influence on health-oriented applications. In this paper, we focus on the task of cross-modal retrieval between food images and cooking recipes. We present Modality-Consistent Embedding Network (MCEN) that learns modality-invariant representations by projecting images and texts to the same embedding space. To capture the latent alignments between modalities, we incorporate stochastic latent variables to explicitly exploit the interactions between textual and visual features. Importantly, our method learns the cross-modal alignments during training but computes embeddings of different modalities independently at inference time for the sake of efficiency. Extensive experimental results clearly demonstrate that the proposed MCEN outperforms all existing approaches on the benchmark Recipe1M dataset and requires less computational cost.

Recent interest in self-supervised dense tracking has yielded rapid progress, but performance still remains far from supervised methods. We propose a dense tracking model trained on videos without any annotations that surpasses previous self-supervised methods on existing benchmarks by a significant margin (+15%), and achieves performance comparable to supervised methods. In this paper, we first reassess the traditional choices used for self-supervised training and reconstruction loss by conducting thorough experiments that finally elucidate the optimal choices. Second, we further improve on existing methods by augmenting our architecture with a crucial memory component. Third, we benchmark on large-scale semi-supervised video object segmentation (aka. dense tracking), and propose a new metric: generalizability. Our first two contributions yield a self-supervised network that for the first time is competitive with supervised methods on standard evaluation metrics of dense tracking. When measuring generalizability, we show self-supervised approaches are actually superior to the majority of supervised methods. We believe this new generalizability metric can better capture the real-world use-cases for dense tracking, and will spur new interest in this research direction.

The ambiguity in image matching is one of main factors decreasing the quality of the 3D model reconstructed by PatchMatch based multiple view stereo. In this paper, we present a novel method, matching ambiguity reduced multiple view stereo (MARMVS) to address this issue. The MARMVS handles the ambiguity in image matching process with three newly proposed strategies: 1) The matching ambiguity is measured by the differential geometry property of image surface with epipolar constraint, which is used as a critical criterion for optimal scale selection of every single pixel with corresponding neighbouring images. 2) The depth of every pixel is initialized to be more close to the true depth by utilizing the depths of its surrounding sparse feature points, which yields faster convergency speed in the following PatchMatch stereo and alleviates the ambiguity introduced by self similar structures of the image. 3) In the last propagation of the PatchMatch stereo, higher priorities are given to those planes with the related 2D image patch possesses less ambiguity, this strategy further propagates a correctly reconstructed surface to raw texture regions. In addition, the proposed method is very efficient even running on consumer grade CPUs, due to proper parameterization and discretization in the depth map computation step. The MARMVS is validated on public benchmarks, and experimental results demonstrate competing performance against the state of the art.

Autonomous vehicles are expected to drive in complex scenarios with several independent non cooperating agents. Path planning for safely navigating in such environments can not just rely on perceiving present location and motion of other agents. It requires instead to predict such variables in a far enough future. In this paper we address the problem of multimodal trajectory prediction exploiting a Memory Augmented Neural Network. Our method learns past and future trajectory embeddings using recurrent neural networks and exploits an associative external memory to store and retrieve such embeddings. Trajectory prediction is then performed by decoding in-memory future encodings conditioned with the observed past. We incorporate scene knowledge in the decoding state by learning a CNN on top of semantic scene maps. Memory growth is limited by learning a writing controller based on the predictive capability of existing embeddings. We show that our method is able to natively perform multi-modal trajectory prediction obtaining state-of-the art results on three datasets. Moreover, thanks to the non-parametric nature of the memory module, we show how once trained our system can continuously improve by ingesting novel patterns.

In most real-world scenarios, labeled training datasets are highly class-imbalanced, where deep neural networks suffer from generalizing to a balanced testing criterion. In this paper, we explore a novel yet simple way to alleviate this issue by augmenting less-frequent classes via translating samples (e.g., images) from more-frequent classes. This simple approach enables a classifier to learn more generalizable features of minority classes, by transferring and leveraging the diversity of the majority information. Our experimental results on a variety of class-imbalanced datasets show that the proposed method improves the generalization on minority classes significantly compared to other existing re-sampling or re-weighting methods. The performance of our method even surpasses those of previous state-of-the-art methods for the imbalanced classification.

We propose an end-to-end learned video compression scheme for low-latency scenarios. Previous methods are limited in using the previous one frame as reference. Our method introduces the usage of the previous multiple frames as references. In our scheme, the motion vector (MV) field is calculated between the current frame and the previous one. With multiple reference frames and associated multiple MV fields, our designed network can generate more accurate prediction of the current frame, yielding less residual. Multiple reference frames also help generate MV prediction, which reduces the coding cost of MV field. We use two deep auto-encoders to compress the residual and the MV, respectively. To compensate for the compression error of the auto-encoders, we further design a MV refinement network and a residual refinement network, taking use of the multiple reference frames as well. All the modules in our scheme are jointly optimized through a single rate-distortion loss function. We use a step-by-step training strategy to optimize the entire scheme. Experimental results show that the proposed method outperforms the existing learned video compression methods for low-latency mode. Our method also performs better than H.265 in both PSNR and MS-SSIM. Our code and models are publicly available.

Neural net compression can be achieved by approximating each layer's weight matrix by a low-rank matrix. The real difficulty in doing this is not in training the resulting neural net (made up of one low-rank matrix per layer), but in determining what the optimal rank of each layer is--effectively, an architecture search problem with one hyperparameter per layer. We show that, with a suitable formulation, this problem is amenable to a mixed discrete-continuous optimization jointly over the ranks and over the matrix elements, and give a corresponding algorithm. We show that this indeed can select ranks much better than existing approaches, making low-rank compression much more attractive than previously thought. For example, we can make a VGG network faster than a ResNet and with nearly the same classification error.

Most object recognition approaches predominantly focus on learning discriminative visual patterns, while overlooking the holistic object structure. Though important, structure modeling usually requires significant manual annotations and therefore is labor-intensive. In this paper, we propose to "look into object" (explicitly yet intrinsically model the object structure) through incorporating self-supervisions into the traditional framework. We show the recognition backbone can be substantially enhanced for more robust representation learning, without any cost of extra annotation and inference speed. Specifically, we first propose an object-extent learning module for localizing the object according to the visual patterns shared among the instances in the same category. We then design a spatial context learning module for modeling the internal structures of the object, through predicting the relative positions within the extent. These two modules can be easily plugged into any backbone networks during training and detached at inference time. Extensive experiments show that our look-into-object approach (LIO) achieves large performance gain on a number of benchmarks, including generic object recognition (ImageNet) and fine-grained object recognition tasks (CUB, Cars, Aircraft). We also show that this learning paradigm is highly generalizable to other tasks such as object detection and segmentation (MS COCO). Project page: https://github.com/JDAI-CV/LIO.

This paper addresses the problem of text-to-video temporal grounding, which aims to identify the time interval in a video semantically relevant to a text query. We tackle this problem using a novel regression-based model that learns to extract a collection of mid-level features for semantic phrases in a text query, which corresponds to important semantic entities described in the query (e.g., actors, objects, and actions), and reflect bi-modal interactions between the linguistic features of the query and the visual features of the video in multiple levels. The proposed method effectively predicts the target time interval by exploiting contextual information from local to global during bi-modal interactions. Through in-depth ablation studies, we find out that incorporating both local and global context in video and text interactions is crucial to the accurate grounding. Our experiment shows that the proposed method outperforms the state of the arts on Charades-STA and ActivityNet Captions datasets by large margins, 7.44% and 4.61% points at Recall@tIoU=0.5 metric, respectively.

We propose a new formulation to non-rigid structure-from-motion that only requires the deforming surface to preserve its differential structure. This is a much weaker assumption than the traditional ones of isometry or conformality. We show that it is nevertheless sufficient to establish local correspondences between the surface in two different images and therefore to perform point-wise reconstruction using only first-order derivatives. To this end, we formulate differential constraints and solve them algebraically using the theory of resultants. We will demonstrate that our approach is more widely applicable, more stable in noisy and sparse imaging conditions and much faster than earlier ones, while delivering similar accuracy. The code is available at https://github.com/cvlab-epfl/diff-nrsfm/.

Shape priors learned from data are commonly used to reconstruct 3D objects from partial or noisy data. Yet no such shape priors are available for indoor scenes, since typical 3D autoencoders cannot handle their scale, complexity, or diversity. In this paper, we introduce Local Implicit Grid Representations, a new 3D shape representation designed for scalability and generality. The motivating idea is that most 3D surfaces share geometric details at some scale -- i.e., at a scale smaller than an entire object and larger than a small patch. We train an autoencoder to learn an embedding of local crops of 3D shapes at that size. Then, we use the decoder as a component in a shape optimization that solves for a set of latent codes on a regular grid of overlapping crops such that an interpolation of the decoded local shapes matches a partial or noisy observation. We demonstrate the value of this proposed approach for 3D surface reconstruction from sparse point observations, showing significantly better results than alternative approaches.

In this paper, we address the task of semantic-guided scene generation. One open challenge widely observed in global image-level generation methods is the difficulty of generating small objects and detailed local texture. To tackle this issue, in this work we consider learning the scene generation in a local context, and correspondingly design a local class-specific generative network with semantic maps as a guidance, which separately constructs and learns sub-generators concentrating on the generation of different classes, and is able to provide more scene details. To learn more discriminative class-specific feature representations for the local generation, a novel classification module is also proposed. To combine the advantage of both global image-level and the local class-specific generation, a joint generation network is designed with an attention fusion module and a dual-discriminator structure embedded. Extensive experiments on two scene image generation tasks show superior generation performance of the proposed model. State-of-the-art results are established by large margins on both tasks and on challenging public benchmarks. The source code and trained models are available at https://github.com/Ha0Tang/LGGAN.