CVPR 2019 Papers

This paper presents a normalization mechanism called Instance-Level Meta Normalization (ILM Norm) to address a learning-to-normalize problem. ILM Norm learns to predict the normalization parameters via both the feature feed-forward and the gradient back-propagation paths. ILM Norm provides a meta normalization mechanism and has several good properties. It can be easily plugged into existing instance-level normalization schemes such as Instance Normalization, Layer Normalization, or Group Normalization. ILM Norm normalizes each instance individually and therefore maintains high performance even when small mini-batch is used. The experimental results show that ILM Norm well adapts to different network architectures and tasks, and it consistently improves the performance of the original models.

Current state-of-the-art instance segmentation methods are not suited for real-time applications like autonomous driving, which require fast execution times at high accuracy. Although the currently dominant proposal-based methods have high accuracy, they are slow and generate masks at a fixed and low resolution. Proposal-free methods, by contrast, can generate masks at high resolution and are often faster, but fail to reach the same accuracy as the proposal-based methods. In this work we propose a new clustering loss function for proposal-free instance segmentation. The loss function pulls the spatial embeddings of pixels belonging to the same instance together and jointly learns an instance-specific clustering bandwidth, maximizing the intersection-over-union of the resulting instance mask. When combined with a fast architecture, the network can perform instance segmentation in real-time while maintaining a high accuracy. We evaluate our method on the challenging Cityscapes benchmark and achieve top results (5% improvement over Mask R-CNN) at more than 10 fps on 2MP images.

In this paper, we introduce a new problem of manipulating a given video by inserting other videos into it. Our main task is, given an object video and a scene video, to insert the object video at a user-specified location in the scene video so that the resulting video looks realistic. We aim to handle different object motions and complex backgrounds without expensive segmentation annotations. As it is difficult to collect training pairs for this problem, we synthesize fake training pairs that can provide helpful supervisory signals when training a neural network with unpaired real data. The proposed network architecture can take both real and fake pairs as input and perform both supervised and unsupervised training in an adversarial learning scheme. To synthesize a realistic video, the network renders each frame based on the current input and previous frames. Within this framework, we observe that injecting noise into previous frames while generating the current frame stabilizes training. We conduct experiments on real-world videos in object tracking and person re-identification benchmark datasets. Experimental results demonstrate that the proposed algorithm is able to synthesize long sequences of realistic videos with a given object video inserted.

Capturing the interesting components of an image is a key aspect of image understanding. When a speaker annotates an image, selecting labels that are informative greatly depends on the prior knowledge of a prospective listener. Motivated by cognitive theories of categorization and communication, we present a new unsupervised approach to model this prior knowledge and quantify the informativeness of a description. Specifically, we compute how knowledge of a label reduces uncertainty over the space of labels and use this uncertainty reduction to rank candidate labels for describing an image. While the full estimation problem is intractable, we describe an efficient algorithm to approximate entropy reduction using a tree-structured graphical model. We evaluate our approach on the open-images dataset using a new evaluation set of 10K ground-truth ratings and find that it achieves over 65% agreement with human raters, close to the upper bound of inter-rater agreement and largely outperforming other unsupervised baseline approaches.

Though image-to-sequence generation models have become overwhelmingly popular in human-computer communications, they suffer from strongly favoring safe generic questions ("What is in this picture?"). Generating uninformative but relevant questions is not sufficient or useful. We argue that a good question is one that has a tightly focused purpose --- one that is aimed at expecting a specific type of response. We build a model that maximizes mutual information between the image, the expected answer and the generated question. To overcome the non-differentiability of discrete natural language tokens, we introduce a variational continuous latent space onto which the expected answers project. We regularize this latent space with a second latent space that ensures clustering of similar answers. Even when we don't know the expected answer, this second latent space can generate goal-driven questions specifically aimed at extracting objects ("what is the person throwing"), attributes, ("What kind of shirt is the person wearing?"), color ("what color is the frisbee?"), material ("What material is the frisbee?"), etc. We quantitatively show that our model is able to retain information about an expected answer category, resulting in more diverse, goal-driven questions. We launch our model on a set of real world images and extract previously unseen visual concepts.

In this work, we present CRIB (Continual Recognition Inspired by Babies), a synthetic incremental object learning environment that can produce data that models visual imagery produced by object exploration in early infancy. CRIB is coupled with a new 3D object dataset, Toys-200, that contains 200 unique toy-like object instances, and is also compatible with existing 3D datasets. Through extensive empirical evaluation of state-of-the-art incremental learning algorithms, we find the novel empirical result that repetition can significantly ameliorate the effects of catastrophic forgetting. Furthermore, we find that in certain cases repetition allows for performance approaching that of batch learning algorithms. Finally, we propose an unsupervised incremental learning task with intriguing baseline results.

Convolutional Neural Network based approaches for monocular 3D human pose estimation usually require a large amount of training images with 3D pose annotations. While it is feasible to provide 2D joint annotations for large corpora of in-the-wild images with humans, providing accurate 3D annotations to such in-the-wild corpora is hardly feasible in practice. Most existing 3D labelled data sets are either synthetically created or feature in-studio images. 3D pose estimation algorithms trained on such data often have limited ability to generalize to real world scene diversity. We therefore propose a new deep learning based method for monocular 3D human pose estimation that shows high accuracy and generalizes better to in-the-wild scenes. It has a network architecture that comprises a new disentangled hidden space encoding of explicit 2D and 3D features, and uses supervision by a new learned projection model from predicted 3D pose. Our algorithm can be jointly trained on image data with 3D labels and image data with only 2D labels. It achieves state-of-the-art accuracy on challenging in-the-wild data.

Recent success of semantic segmentation approaches on demanding road driving datasets has spurred interest in many related application fields. Many of these applications involve real-time prediction on mobile platforms such as cars, drones and various kinds of robots. Real-time setup is challenging due to extraordinary computational complexity involved. Many previous works address the challenge with custom lightweight architectures which decrease computational complexity by reducing depth, width and layer capacity with respect to general purpose architectures. We propose an alternative approach which achieves a significantly better performance across a wide range of computing budgets. First, we rely on a light-weight general purpose architecture as the main recognition engine. Then, we leverage light-weight upsampling with lateral connections as the most cost-effective solution to restore the prediction resolution. Finally, we propose to enlarge the receptive field by fusing shared features at multiple resolutions in a novel fashion. Experiments on several road driving datasets show a substantial advantage of the proposed approach, either with ImageNet pre-trained parameters or when we learn from scratch. Our Cityscapes test submission entitled SwiftNetRN-18 delivers 75.5% MIoU and achieves 39.9 Hz on 1024x2048 images on GTX1080Ti.

We present an efficient approach for leveraging the knowledge from multiple modalities in training unimodal 3D convolutional neural networks (3D-CNNs) for the task of dynamic hand gesture recognition. Instead of explicitly combining multimodal information, which is commonplace in many state-of-the-art methods, we propose a different framework in which we embed the knowledge of multiple modalities in individual networks so that each unimodal network can achieve an improved performance. In particular, we dedicate separate networks per available modality and enforce them to collaborate and learn to develop networks with common semantics and better representations. We introduce a "spatiotemporal semantic alignment" loss (SSA) to align the content of the features from different networks. In addition, we regularize this loss with our proposed "focal regularization parameter" to avoid negative knowledge transfer. Experimental results show that our framework improves the test time recognition accuracy of unimodal networks, and provides the state-of-the-art performance on various dynamic hand gesture recognition datasets.

Though CNNs have achieved the state-of-the-art performance on various vision tasks, they are vulnerable to adversarial examples --- crafted by adding human-imperceptible perturbations to clean images. However, most of the existing adversarial attacks only achieve relatively low success rates under the challenging black-box setting, where the attackers have no knowledge of the model structure and parameters. To this end, we propose to improve the transferability of adversarial examples by creating diverse input patterns. Instead of only using the original images to generate adversarial examples, our method applies random transformations to the input images at each iteration. Extensive experiments on ImageNet show that the proposed attack method can generate adversarial examples that transfer much better to different networks than existing baselines. By evaluating our method against top defense solutions and official baselines from NIPS 2017 adversarial competition, the enhanced attack reaches an average success rate of 73.0%, which outperforms the top-1 attack submission in the NIPS competition by a large margin of 6.6%. We hope that our proposed attack strategy can serve as a strong benchmark baseline for evaluating the robustness of networks to adversaries and the effectiveness of different defense methods in the future. Code is available at https://github.com/cihangxie/DI-2-FGSM.

Semantic segmentation requires large amounts of pixel-wise annotations to learn accurate models. In this paper, we present a video prediction-based methodology to scale up training sets by synthesizing new training samples in order to improve the accuracy of semantic segmentation networks. We exploit video prediction models' ability to predict future frames in order to also predict future labels. A joint propagation strategy is also proposed to alleviate mis-alignments in synthesized samples. We demonstrate that training segmentation models on datasets augmented by the synthesized samples leads to significant improvements in accuracy. Furthermore, we introduce a novel boundary label relaxation technique that makes training robust to annotation noise and propagation artifacts along object boundaries. Our proposed methods achieve state-of-the-art mIoUs of 83.5% on Cityscapes and 82.9% on CamVid. Our single model, without model ensembles, achieves 72.8% mIoU on the KITTI semantic segmentation test set, which surpasses the winning entry of the ROB challenge 2018.

Referring expression grounding aims at locating certain objects or persons in an image with a referring expression, where the key challenge is to comprehend and align various types of information from visual and textual domain, such as visual attributes, location and interactions with surrounding regions. Although the attention mechanism has been successfully applied for cross-modal alignments, previous attention models focus on only the most dominant features of both modalities, and neglect the fact that there could be multiple comprehensive textual-visual correspondences between images and referring expressions. To tackle this issue, we design a novel cross-modal attention-guided erasing approach, where we discard the most dominant information from either textual or visual domains to generate difficult training samples online, and to drive the model to discover complementary textual-visual correspondences. Extensive experiments demonstrate the effectiveness of our proposed method, which achieves state-of-the-art performance on three referring expression grounding datasets.

As an indicator of human attention gaze is a subtle behavioral cue which can be exploited in many applications. However, inferring 3D gaze direction is challenging even for deep neural networks given the lack of large amount of data (groundtruthing gaze is expensive and existing datasets use different setups) and the inherent presence of gaze biases due to person-specific difference. In this work, we address the problem of person-specific gaze model adaptation from only a few reference training samples. The main and novel idea is to improve gaze adaptation by generating additional training samples through the synthesis of gaze-redirected eye images from existing reference samples. In doing so, our contributions are threefold:(i) we design our gaze redirection framework from synthetic data, allowing us to benefit from aligned training sample pairs to predict accurate inverse mapping fields; (ii) we proposed a self-supervised approach for domain adaptation; (iii) we exploit the gaze redirection to improve the performance of person-specific gaze estimation. Extensive experiments on two public datasets demonstrate the validity of our gaze retargeting and gaze estimation framework.

Spatio-temporal action localization consists of three levels of tasks: spatial localization, action classification, and temporal segmentation. In this work, we propose a new Progressive Cross-stream Cooperation (PCSC) framework to iterative improve action localization results and generate better bounding boxes for one stream (i.e., Flow/RGB) by leveraging both region proposals and features from another stream (i.e., RGB/Flow) in an iterative fashion. Specifically, we first generate a larger set of region proposals by combining the latest region proposals from both streams, from which we can readily obtain a larger set of labelled training samples to help learn better action detection models. Second, we also propose a new message passing approach to pass information from one stream to another stream in order to learn better representations, which also leads to better action detection models. As a result, our iterative framework progressively improves action localization results at the frame level. To improve action localization results at the video level, we additionally propose a new strategy to train class-specific actionness detectors for better temporal segmentation, which can be readily learnt by using the training samples around temporal boundaries. Comprehensive experiments on two benchmark datasets UCF-101-24 and J-HMDB demonstrate the effectiveness of our newly proposed approaches for spatio-temporal action localization in realistic scenarios.

Road network extraction from satellite images often produce fragmented road segments leading to road maps unfit for real applications. Pixel-wise classification fails to predict topologically correct and connected road masks due to the absence of connectivity supervision and difficulty in enforcing topological constraints. In this paper, we propose a connectivity task called Orientation Learning, motivated by the human behavior of annotating roads by tracing it at a specific orientation. We also develop a stacked multi-branch convolutional module to effectively utilize the mutual information between orientation learning and segmentation tasks. These contributions ensure that the model predicts topologically correct and connected road masks. We also propose Connectivity Refinement approach to further enhance the estimated road networks. The refinement model is pre-trained to connect and refine the corrupted ground-truth masks and later fine-tuned to enhance the predicted road masks. We demonstrate the advantages of our approach on two diverse road extraction datasets SpaceNet and DeepGlobe. Our approach improves over the state-of-the-art techniques by 9% and 7.5% in road topology metric on SpaceNet and DeepGlobe, respectively.

Structural pruning of neural network parameters reduces computational, energy, and memory transfer costs during inference. We propose a novel method that estimates the contribution of a neuron (filter) to the final loss and iteratively removes those with smaller scores. We describe two variations of our method using the first and second-order Taylor expansions to approximate a filter's contribution. Both methods scale consistently across any network layer without requiring per-layer sensitivity analysis and can be applied to any kind of layer, including skip connections. For modern networks trained on ImageNet, we measured experimentally a high (>93%) correlation between the contribution computed by our methods and a reliable estimate of the true importance. Pruning with the proposed methods led to an improvement over state-of-the-art in terms of accuracy, FLOPs, and parameter reduction. On ResNet-101, we achieve a 40% FLOPS reduction by removing 30% of the parameters, with a loss of 0.02% in the top-1 accuracy on ImageNet.

Recently, unsupervised exemplar-based image-to-image translation, conditioned on a given exemplar without the paired data, has accomplished substantial advancements. In order to transfer the information from an exemplar to an input image, existing methods often use a normalization technique, e.g., adaptive instance normalization, that controls the channel-wise statistics of an input activation map at a particular layer, such as the mean and the variance. Meanwhile, style transfer approaches similar task to image translation by nature, demonstrated superior performance by using the higher-order statistics such as covariance among channels in representing a style. In detail, it works via whitening (given a zero-mean input feature, transforming its covariance matrix into the identity). followed by coloring (changing the covariance matrix of the whitened feature to those of the style feature). However, applying this approach in image translation is computationally intensive and error-prone due to the expensive time complexity and its non-trivial backpropagation. In response, this paper proposes an end-to-end approach tailored for image translation that efficiently approximates this transformation with our novel regularization methods. We further extend our approach to a group-wise form for memory and time efficiency as well as image quality. Extensive qualitative and quantitative experiments demonstrate that our proposed method is fast, both in training and inference, and highly effective in reflecting the style of an exemplar.

The image, question (combined with the history for de-referencing), and the corresponding answer are three vital components of visual dialog. Classical visual dialog systems integrate the image, question, and history to search for or generate the best matched answer, and so, this approach significantly ignores the role of the answer. In this paper, we devise a novel image-question-answer synergistic network to value the role of the answer for precise visual dialog. We extend the traditional one-stage solution to a two-stage solution. In the first stage, candidate answers are coarsely scored according to their relevance to the image and question pair. Afterward, in the second stage, answers with high probability of being correct are re-ranked by synergizing with image and question. On the Visual Dialog v1.0 dataset, the proposed synergistic network boosts the discriminative visual dialog model to achieve a new state-of-the-art of 57.88% normalized discounted cumulative gain. A generative visual dialog model equipped with the proposed technique also shows promising improvements.

Due to the significant information loss in low-resolution (LR) images, it has become extremely challenging to further advance the state-of-the-art of single image super-resolution (SISR). Reference-based super-resolution (RefSR), on the other hand, has proven to be promising in recovering high-resolution (HR) details when a reference (Ref) image with similar content as that of the LR input is given. However, the quality of RefSR can degrade severely when Ref is less similar. This paper aims to unleash the potential of RefSR by leveraging more texture details from Ref images with stronger robustness even when irrelevant Ref images are provided. Inspired by the recent work on image stylization, we formulate the RefSR problem as neural texture transfer. We design an end-to-end deep model which enriches HR details by adaptively transferring the texture from Ref images according to their textural similarity. Instead of matching content in the raw pixel space as done by previous methods, our key contribution is a multi-level matching conducted in the neural space. This matching scheme facilitates multi-scale neural transfer that allows the model to benefit more from those semantically related Ref patches, and gracefully degrade to SISR performance on the least relevant Ref inputs. We build a benchmark dataset for the general research of RefSR, which contains Ref images paired with LR inputs with varying levels of similarity. Both quantitative and qualitative evaluations demonstrate the superiority of our method over state-of-the-art.