CVPR 2019 Papers

In this work, we present a novel meta-learning algorithm that regresses model parameters for novel tasks for which no ground truth is available (zero-shot tasks). In order to adapt to novel zero-shot tasks, our meta-learner learns from the model parameters of known tasks (with ground truth) and the correlation of known tasks to zero-shot tasks. Such intuition finds its foothold in cognitive science, where a subject (human baby) can adapt to a novel concept (depth understanding) by correlating it with old concepts (hand movement or self-motion), without receiving an explicit supervision. We evaluated our model on the Taskonomy dataset, with four tasks as zero-shot: surface normal, room layout, depth and camera pose estimation. These tasks were chosen based on the data acquisition complexity and the complexity associated with the learning process using a deep network. Our proposed methodolgy outperforms state-of-the-art models (which use ground truth) on each of our zero-shot tasks, showing promise on zero-shot task transfer. We also conducted extensive experiments to study the various choices of our methodology, as well as showed how the proposed method can also be used in transfer learning. To the best of our knowledge, this is the first such effort on zero-shot learning in the task space.

Classifiers used in the wild, in particular for safety-critical systems, should not only have good generalization properties but also should know when they don't know, in particular make low confidence predictions far away from the training data. We show that ReLU type neural networks which yield a piecewise linear classifier function fail in this regard as they produce almost always high confidence predictions far away from the training data. For bounded domains like images we propose a new robust optimization technique similar to adversarial training which enforces low confidence predictions far away from the training data. We show that this technique is surprisingly effective in reducing the confidence of predictions far away from the training data while maintaining high confidence predictions and test error on the original classification task compared to standard training.

Path forecasting is a pivotal step toward understanding dynamic scenes and an emerging topic in the computer vi- sion field. This task is challenging due to the multimodal nature of the future, namely, given a partial history, there is more than one plausible prediction. Yet, the state-of-the-art methods seem not fully responsive to this innate variabil- ity. Hence, how to better foresee the forthcoming trajectory in dynamic scenes has to be more thoroughly pursued. To this end, we propose a novel Imitative Decision Learning (IDL) approach. It delves deeper into the key that inher- ently characterizes the multimodality - the latent decision. The proposed IDL first infers the distribution of such latent decisions by learning from moving histories. A policy is then generated by taking the sampled latent decision into account to predict the future. Different plausible upcoming paths corresponds to each sampled latent decision. This ap- proach significantly differs from the mainstream literature that relies on a predefined latent variable to extrapolate di- verse predictions. In order to augment the understanding of the latent decision and resultant mutimodal future, we in- vestigate their connection through mutual information op- timization. Moreover, the proposed IDL integrates spatial and temporal dependencies into one single framework, in contrast to handling them with two-step settings. As a re- sult, our approach enables simultaneous anticipation of the paths of all pedestrians in the scene. We assess our pro- posal on the large-scale SAP, ETH and UCY datasets. The experiments show that IDL introduces considerable margin improvements with respect to recent leading studies.

Can performance on the task of action quality assessment (AQA) be improved by exploiting a description of the action and its quality? Current AQA and skills assessment approaches propose to learn features that serve only one task - estimating the final score. In this paper, we propose to learn spatio-temporal features that explain three related tasks - fine-grained action recognition, commentary generation, and estimating the AQA score. A new multitask-AQA dataset, the largest to date, comprising of 1412 diving samples was collected to evaluate our approach (http://rtis.oit.unlv.edu/datasets.html). We show that our MTL approach outperforms STL approach using two different kinds of architectures: C3D-AVG and MSCADC. The C3D-AVG-MTL approach achieves the new state-of-the-art performance with a rank correlation of 90.44%. Detailed experiments were performed to show that MTL offers better generalization than STL, and representations from action recognition models are not sufficient for the AQA task and instead should be learned.

In two-view geometry, camera models and motion types are used as key knowledge along with the image point correspondences in order to solve several key problems of 3D vision. Problems such as Structure-from-Motion (SfM) and camera self-calibration are tackled under the assumptions of a specific camera projection model and motion type. However, these key assumptions may not be always justified, i.e.., we may often know neither the camera model nor the motion type beforehand. In that context, one can extract only the point correspondences between images. From such correspondences, recovering two-view relationship --expressed by the unknown camera model and motion type-- remains to be an unsolved problem. In this paper, we tackle this problem in two steps. First, we propose a method that computes the correct two-view relationship in the presence of noise and outliers. Later, we study different possibilities to disambiguate the obtained relationships into camera model and motion type. By extensive experiments on both synthetic and real data, we verify our theory and assumptions in practical settings.

Recent studies have shown remarkable advances in 3D human pose estimation from monocular images, with the help of large-scale in-door 3D datasets and sophisticated network architectures. However, the generalizability to different environments remains an elusive goal. In this work, we propose a geometry-aware 3D representation for the human pose to address this limitation by using multiple views in a simple auto-encoder model at the training stage and only 2D keypoint information as supervision. A view synthesis framework is proposed to learn the shared 3D representation between viewpoints with synthe- sizing the human pose from one viewpoint to the other one. Instead of performing a direct transfer in the raw image- level, we propose a skeleton-based encoder-decoder mechanism to distil only pose-related representation in the latent space. A learning-based representation consistency constraint is further introduced to facilitate the robustness of latent 3D representation. Since the learnt representation encodes 3D geometry information, mapping it to 3D pose will be much easier than conventional frameworks that use an image or 2D coordinates as the input of 3D pose estimator. We demonstrate our approach on the task of 3D human pose estimation. Comprehensive experiments on three popular benchmarks show that our model can significantly improve the performance of state-of-the-art methods with simply injecting the representation as a robust 3D prior.

This paper presents a novel approach for learning instance segmentation with image-level class labels as supervision. Our approach generates pseudo instance segmentation labels of training images, which are used to train a fully supervised model. For generating the pseudo labels, we first identify confident seed areas of object classes from attention maps of an image classification model, and propagate them to discover the entire instance areas with accurate boundaries. To this end, we propose IRNet, which estimates rough areas of individual instances and detects boundaries between different object classes. It thus enables to assign instance labels to the seeds and to propagate them within the boundaries so that the entire areas of instances can be estimated accurately. Furthermore, IRNet is trained with inter-pixel relations on the attention maps, thus no extra supervision is required. Our method with IRNet achieves an outstanding performance on the PASCAL VOC 2012 dataset, surpassing not only previous state-of-the-art trained with the same level of supervision, but also some of previous models relying on stronger supervision.

We propose, WarpGAN, a fully automatic network that can generate caricatures given an input face photo. Besides transferring rich texture styles, WarpGAN learns to automatically predict a set of control points that can warp the photo into a caricature, while preserving identity. We introduce an identity-preserving adversarial loss that aids the discriminator to distinguish between different subjects. Moreover, WarpGAN allows customization of the generated caricatures by controlling the exaggeration extent and the visual styles. Experimental results on a public domain dataset, WebCaricature, show that WarpGAN is capable of generating caricatures that not only preserve the identities but also outputs a diverse set of caricatures for each input photo. Five caricature experts suggest that caricatures generated by WarpGAN are visually similar to hand-drawn ones and only prominent facial features are exaggerated.

In this work, we propose a novel adaptive spatially-regularized correlation filters (ASRCF) model to simultaneously optimize the filter coefficients and the spatial regularization weight. First, this adaptive spatial regularization scheme could learn an effective spatial weight for a specific object and its appearance variations, and therefore result in more reliable filter coefficients during the tracking process. Second, our ASRCF model can be effectively optimized based on the alternating direction method of multipliers, where each subproblem has the closed-from solution. Third, our tracker applies two kinds of CF models to estimate the location and scale respectively. The location CF model exploits ensembles of shallow and deep features to determine the optimal position accurately. The scale CF model works on multi-scale shallow features to estimate the optimal scale efficiently. Extensive experiments on five recent benchmarks show that our tracker performs favorably against many state-of-the-art algorithms, with real-time performance of 28fps.

We introduce the Action Transformer model for recognizing and localizing human actions in video clips. We repurpose a Transformer-style architecture to aggregate features from the spatiotemporal context around the person whose actions we are trying to classify. We show that by using high-resolution, person-specific, class-agnostic queries, the model spontaneously learns to track individual people and to pick up on semantic context from the actions of others. Additionally its attention mechanism learns to emphasize hands and faces, which are often crucial to discriminate an action - all without explicit supervision other than boxes and class labels. We train and test our Action Transformer network on the Atomic Visual Actions (AVA) dataset, outperforming the state-of-the-art by a significant margin using only raw RGB frames as input.

For unsupervised domain adaptation, the target domain error can be provably reduced by having a shared input representation that makes the source and target domains indistinguishable from each other. Very recently it has been shown that it is not only critical to match the marginal input distributions, but also align the output class distributions. The latter can be achieved by minimizing the maximum discrepancy of predictors. In this paper, we take this principle further by proposing a more systematic and effective way to achieve hypothesis consistency using Gaussian processes (GP). The GP allows us to induce a hypothesis space of classifiers from the posterior distribution of the latent random functions, turning the learning into a large-margin posterior separation problem, significantly easier to solve than previous approaches based on adversarial minimax optimization. We formulate a learning objective that effectively influences the posterior to minimize the maximum discrepancy. This is shown to be equivalent to maximizing margins and minimizing uncertainty of the class predictions in the target domain. Empirical results demonstrate that our approach leads to state-to-the-art performance superior to existing methods on several challenging benchmarks for domain adaptation.

Although unsupervised person re-identification (RE-ID) has drawn increasing research attentions due to its potential to address the scalability problem of supervised RE-ID models, it is very challenging to learn discriminative information in the absence of pairwise labels across disjoint camera views. To overcome this problem, we propose a deep model for the soft multilabel learning for unsupervised RE-ID. The idea is to learn a soft multilabel (real-valued label likelihood vector) for each unlabeled person by comparing the unlabeled person with a set of known reference persons from an auxiliary domain. We propose the soft multilabel-guided hard negative mining to learn a discriminative embedding for the unlabeled target domain by exploring the similarity consistency of the visual features and the soft multilabels of unlabeled target pairs. Since most target pairs are cross-view pairs, we develop the cross-view consistent soft multilabel learning to achieve the learning goal that the soft multilabels are consistently good across different camera views. To enable effecient soft multilabel learning, we introduce the reference agent learning to represent each reference person by a reference agent in a joint embedding. We evaluate our unified deep model on Market-1501 and DukeMTMC-reID. Our model outperforms the state-of-the-art unsupervised RE-ID methods by clear margins. Code is available at https://github.com/KovenYu/MAR.

In this paper, we address unsupervised pose-guided person image generation, which is known challenging due to non-rigid deformation. Unlike previous methods learning a rock-hard direct mapping between human bodies, we propose a new pathway to decompose the hard mapping into two more accessible subtasks, namely, semantic parsing transformation and appearance generation. Firstly, a semantic generative network is proposed to transform between semantic parsing maps, in order to simplify the non-rigid deformation learning. Secondly, an appearance generative network learns to synthesize semantic-aware textures. Thirdly, we demonstrate that training our framework in an end-to-end manner further refines the semantic maps and final results accordingly. Our method is generalizable to other semantic-aware person image generation tasks, e.g., clothing texture transfer and controlled image manipulation. Experimental results demonstrate the superiority of our method on DeepFashion and Market-1501 datasets, especially in keeping the clothing attributes and better body shapes.

Large intra-class variation is the result of changes in multiple object characteristics. Images, however, only show the superposition of different variable factors such as appearance or shape. Therefore, learning to disentangle and represent these different characteristics poses a great challenge, especially in the unsupervised case. Moreover, large object articulation calls for a flexible part-based model. We present an unsupervised approach for disentangling appearance and shape by learning parts consistently over all instances of a category. Our model for learning an object representation is trained by simultaneously exploiting invariance and equivariance constraints between synthetically transformed images. Since no part annotation or prior information on an object class is required, the approach is applicable to arbitrary classes. We evaluate our approach on a wide range of object categories and diverse tasks including pose prediction, disentangled image synthesis, and video-to-video translation. The approach outperforms the state-of-the-art on unsupervised keypoint prediction and compares favorably even against supervised approaches on the task of shape and appearance transfer.

We introduce the first completely unsupervised correspondence learning approach for deformable 3D shapes. Key to our model is the understanding that natural deformations (such as changes in pose) approximately preserve the metric structure of the surface, yielding a natural criterion to drive the learning process toward distortion-minimizing predictions. On this basis, we overcome the need for annotated data and replace it by a purely geometric criterion. The resulting learning model is class-agnostic, and is able to leverage any type of deformable geometric data for the training phase. In contrast to existing supervised approaches which specialize on the class seen at training time, we demonstrate stronger generalization as well as applicability to a variety of challenging settings. We showcase our method on a wide selection of correspondence benchmarks, where we outperform other methods in terms of accuracy, generalization, and efficiency.

Machine learning techniques work best when the data used for training resembles the data used for evaluation. This holds true for learned single-image denoising algorithms, which are applied to real raw camera sensor readings but, due to practical constraints, are often trained on synthetic image data. Though it is understood that generalizing from synthetic to real images requires careful consideration of the noise properties of camera sensors, the other aspects of an image processing pipeline (such as gain, color correction, and tone mapping) are often overlooked, despite their significant effect on how raw measurements are transformed into finished images. To address this, we present a technique to "unprocess" images by inverting each step of an image processing pipeline, thereby allowing us to synthesize realistic raw sensor measurements from commonly available Internet photos. We additionally model the relevant components of an image processing pipeline when evaluating our loss function, which allows training to be aware of all relevant photometric processing that will occur after denoising. By unprocessing and processing training data and model outputs in this way, we are able to train a simple convolutional neural network that has 14%-38% lower error rates and is 9x-18x faster than the previous state of the art on the Darmstadt Noise Dataset, and generalizes to sensors outside of that dataset as well.

We propose the Unified Visual-Semantic Embeddings (Unified VSE) for learning a joint space of visual representation and textual semantics. The model unifies the embeddings of concepts at different levels: objects, attributes, relations, and full scenes. We view the sentential semantics as a combination of different semantic components such as objects and relations; their embeddings are aligned with different image regions. A contrastive learning approach is proposed for the effective learning of this fine-grained alignment from only image-caption pairs. We also present a simple yet effective approach that enforces the coverage of caption embeddings on the semantic components that appear in the sentence. We demonstrate that the Unified VSE outperforms baselines on cross-modal retrieval tasks; the enforcement of the semantic coverage improves the model's robustness in defending text-domain adversarial attacks. Moreover, our model empowers the use of visual cues to accurately resolve word dependencies in novel sentences.

This paper presents a new neural network for enhancing underexposed photos. Instead of directly learning an image-to-image mapping as previous work, we introduce intermediate illumination in our network to associate the input with expected enhancement result, which augments the network's capability to learn complex photographic adjustment from expert-retouched input/output image pairs. Based on this model, we formulate a loss function that adopts constraints and priors on the illumination, prepare a new dataset of 3,000 underexposed image pairs, and train the network to effectively learn a rich variety of adjustment for diverse lighting conditions. By these means, our network is able to recover clear details, distinct contrast, and natural color in the enhancement results. We perform extensive experiments on the benchmark MIT-Adobe FiveK dataset and our new dataset, and show that our network is effective to deal with previously challenging images.

In this paper, we propose a unified panoptic segmentation network (UPSNet) for tackling the newly proposed panoptic segmentation task. On top of a single backbone residual network, we first design a deformable convolution based semantic segmentation head and a Mask R-CNN style instance segmentation head which solve these two subtasks simultaneously. More importantly, we introduce a parameter-free panoptic head which solves the panoptic segmentation via pixel-wise classification. It first leverages the logits from the previous two heads and then innovatively expands the representation for enabling prediction of an extra unknown class which helps better resolving the conflicts between semantic and instance segmentation. Besides, it handles the challenge caused by the varying number of instances and permits back propagation to the bottom modules in an end-to-end manner. Extensive experimental results on Cityscapes, COCO and our internal dataset demonstrate that our UPSNet achieves state-of-the-art performance with much faster inference. Code has been made available at: https://github.com/uber-research/UPSNet