CVPR 2019 Papers

Existing dichromatic color constancy approach commonly requires a number of spatial pixels which have high specularity. In this paper, we propose a novel approach to estimate the illuminant chromaticity of AC light source using high-speed camera. We found that the temporal observations of an image pixel at a fixed location distribute on an identical dichromatic plane. Instead of spatial pixels with high specularity, multiple temporal samples of a pixel are exploited to determine AC pixels for dichromatic plane estimation, whose pixel intensity is sinusoidally varying well. A dichromatic plane is calculated per each AC pixel, and illuminant chromaticity is determined by the intersection of dichromatic planes. From multiple dichromatic planes, an optimal illuminant is estimated with a novel MAP framework. It is shown that the proposed method outperforms both existing dichromatic based methods and temporal color constancy methods, irrespective of the amount of specularity.

We tackle the problem of semantic boundary prediction, which aims to identify pixels that belong to object(class) boundaries. We notice that relevant datasets consist of a significant level of label noise, reflecting the fact that precise annotations are laborious to get and thus annotators trade-off quality with efficiency. We aim to learn sharp and precise semantic boundaries by explicitly reasoning about annotation noise during training. We propose a simple new layer and loss that can be used with existing learning-based boundary detectors. Our layer/loss enforces the detector to predict a maximum response along the normal direction at an edge, while also regularizing its direction. We further reason about true object boundaries during training using a level set formulation, which allows the network to learn from misaligned labels in an end-to-end fashion. Experiments show that we improve over the CASENet backbone network by more than 4% in terms of MF(ODS) and 18.61% in terms of AP, outperforming all current state-of-the-art methods including those that deal with alignment. Furthermore, we show that our learned network can be used to significantly improve coarse segmentation labels, lending itself as an efficient way to label new data.

Deep Neural Networks (DNNs) have recently led to significant improvements in many fields. However, DNNs are vulnerable to adversarial examples which are samples with imperceptible perturbations while dramatically misleading the DNNs. Moreover, adversarial examples can be used to perform an attack on various kinds of DNN based systems, even if the adversary has no access to the underlying model. Many defense methods have been proposed, such as obfuscating gradients of the networks or detecting adversarial examples. However it is proved out that these defense methods are not effective or cannot resist secondary adversarial attacks. In this paper, we point out that steganalysis can be applied to adversarial examples detection, and propose a method to enhance steganalysis features by estimating the probability of modifications caused by adversarial attacks. Experimental results show that the proposed method can accurately detect adversarial examples. Moreover, secondary adversarial attacks are hard to be directly performed to our method because our method is not based on a neural network but based on high-dimensional artificial features and Fisher Linear Discriminant ensemble.

State of the art deep generative networks have achieved such realism that they can be suspected of memorizing training images. It is why it is not uncommon to include visualizations of training set nearest neighbors, to suggest generated images are not simply memorized. We argue this is not sufficient and motivates studying overfitting of deep generators with more scrutiny. We address this question by i) showing how simple losses are highly effective at reconstructing images for deep generators ii) analyzing the statistics of reconstruction errors for training versus validation images. Using this methodology, we show that pure GAN models appear to generalize well, in contrast with those using hybrid adversarial losses, which are amongst the most widely applied generative methods. We also show that standard GAN evaluation metrics fail to capture memorization for some deep generators. Finally, we note the ramifications of memorization on data privacy. Considering the already widespread application of generative networks, we provide a step in the right direction towards the important yet incomplete picture of generative overfitting.

Retrieving object instances among cluttered scenes efficiently requires compact yet comprehensive regional image representations. Intuitively, object semantics can help build the index that focuses on the most relevant regions. However, due to the lack of bounding-box datasets for objects of interest among retrieval benchmarks, most recent work on regional representations has focused on either uniform or class-agnostic region selection. In this paper, we first fill the void by providing a new dataset of landmark bounding boxes, based on the Google Landmarks dataset, that includes 94k images with manually curated boxes from 15k unique landmarks. Then, we demonstrate how a trained landmark detector, using our new dataset, can be leveraged to index image regions and improve retrieval accuracy while being much more efficient than existing regional methods. In addition, we introduce a novel regional aggregated selective match kernel (R-ASMK) to effectively combine information from detected regions into an improved holistic image representation. R-ASMK boosts image retrieval accuracy substantially with no dimensionality increase, while even outperforming systems that index image regions independently. Our complete image retrieval system improves upon the previous state-of-the-art by significant margins on the Revisited Oxford and Paris datasets. Code and data will be released.

This paper presents a novel framework to recover detailed human body shapes from a single image. It is a challenging task due to factors such as variations in human shapes, body poses, and viewpoints. Prior methods typically attempt to recover the human body shape using a parametric based template that lacks the surface details. As such the resulting body shape appears to be without clothing. In this paper, we propose a novel learning-based framework that combines the robustness of parametric model with the flexibility of free-form 3D deformation. We use the deep neural networks to refine the 3D shape in a Hierarchical Mesh Deformation (HMD) framework, utilizing the constraints from body joints, silhouettes, and per-pixel shading information. We are able to restore detailed human body shapes beyond skinned models. Experiments demonstrate that our method has outperformed previous state-of-the-art approaches, achieving better accuracy in terms of both 2D IoU number and 3D metric distance. The code is available in https://github.com/zhuhao-nju/hmd.git.

Delicate feature representation about object parts plays a critical role in fine-grained recognition. For example, experts can even distinguish fine-grained objects relying only on object parts according to professional knowledge. In this paper, we propose a novel "Destruction and Construction Learning" (DCL) method to enhance the difficulty of fine-grained recognition and exercise the classification model to acquire expert knowledge. Besides the standard classification backbone network, another "destruction and construction" stream is introduced to carefully "destruct" and then "reconstruct" the input image, for learning discriminative regions and features. More specifically, for "destruction", we first partition the input image into local regions and then shuffle them by a Region Confusion Mechanism (RCM). To correctly recognize these destructed images, the classification network has to pay more attention to discriminative regions for spotting the differences. To compensate the noises introduced by RCM, an adversarial loss, which distinguishes original images from destructed ones, is applied to reject noisy patterns introduced by RCM. For "construction", a region alignment network, which tries to restore the original spatial layout of local regions, is followed to model the semantic correlation among local regions. By jointly training with parameter sharing, our proposed DCL injects more discriminative local details to the classification network. Experimental results show that our proposed framework achieves state-of-the-art performance on three standard benchmarks. Moreover, our proposed method does not need any external knowledge during training, and there is no computation overhead at inference time except the standard classification network feed-forwarding. Source code: https://github.com/JDAI-CV/DCL.

Recently, the state-of-the-art models for image captioning have overtaken human performance based on the most popular metrics, such as BLEU, METEOR, ROUGE and CIDEr. Does this mean we have solved the task of image captioning The above metrics only measure the similarity of the generated caption to the human annotations, which reflects its accuracy. However, an image contains many concepts and multiple levels of detail, and thus there is a variety of captions that express different concepts and details that might be interesting for different humans. Therefore only evaluating accuracy is not sufficient for measuring the performance of captioning models --- the diversity of the generated captions should also be considered. In this paper, we proposed a new metric for measuring the diversity of image captions, which is derived from latent semantic analysis and kernelized to use CIDEr similarity. We conduct extensive experiments to re-evaluate recent captioning models in the context of both diversity and accuracy. We find that there is still a large gap between the model and human performance in terms of both accuracy and diversity, and the models that have optimized accuracy (CIDEr) have low diversity. We also show that balancing the cross-entropy loss and CIDEr reward in reinforcement learning during training can effectively control the tradeoff between diversity and accuracy of the generated captions.

Video frame interpolation aims to synthesize nonexistent frames in-between the original frames. While significant advances have been made from the recent deep convolutional neural networks, the quality of interpolation is often reduced due to large object motion or occlusion. In this work, we propose a video frame interpolation method which explicitly detects the occlusion by exploring the depth information. Specifically, we develop a depth-aware flow projection layer to synthesize intermediate flows that preferably sample closer objects than farther ones. In addition, we learn hierarchical features to gather contextual information from neighboring pixels. The proposed model then warps the input frames, depth maps, and contextual features based on the optical flow and local interpolation kernels for synthesizing the output frame. Our model is compact, efficient, and fully differentiable. Quantitative and qualitative results demonstrate that the proposed model performs favorably against state-of-the-art frame interpolation methods on a wide variety of datasets. The source code and pre-trained model are available at https://github.com/baowenbo/DAIN.

Rain is a common weather phenomenon, where object visibility varies with depth from the camera and objects faraway are visually blocked more by fog than by rain streaks. Existing methods and datasets for rain removal, however, ignore these physical properties, thereby limiting the rain removal efficiency on real photos. In this work, we first analyze the visual effects of rain subject to scene depth and formulate a rain imaging model collectively with rain streaks and fog; by then, we prepare a new dataset called RainCityscapes with rain streaks and fog on real outdoor photos. Furthermore, we design an end-to-end deep neural network, where we train it to learn depth-attentional features via a depth-guided attention mechanism, and regress a residual map to produce the rain-free image output. We performed various experiments to visually and quantitatively compare our method with several state-of-the-art methods to demonstrate its superiority over the others.

In this paper, we propose a hybrid depth imaging system in which a polarisation camera is augmented by a second image from a standard digital camera. For this modest increase in equipment complexity over conventional shape-from-polarisation, we obtain a number of benefits that enable us to overcome longstanding problems with the polarisation shape cue. The stereo cue provides a depth map which, although coarse, is metrically accurate. This is used as a guide surface for disambiguation of the polarisation surface normal estimates using a higher order graphical model. In turn, these are used to estimate diffuse albedo. By extending a previous shape-from-polarisation method to the perspective case, we show how to compute dense, detailed maps of absolute depth, while retaining a linear formulation. We show that our hybrid method is able to recover dense 3D geometry that is superior to state-of-the-art shape-from-polarisation or two view stereo alone.

Depth completion involves estimating a dense depth image from sparse depth measurements, often guided by a color image. While linear upsampling is straight forward, it results in depth pixels being interpolated in empty space across discontinuities between objects. Current methods use deep networks to maintain gaps between objects. Nevertheless depth smearing remains a challenge. We propose a new representation for depth called Depth Coefficients (DC) to address this problem. It enables convolutions to more easily avoid inter-object depth mixing. We also show that the standard Mean Squared Error (MSE) loss function can promote depth mixing, and so we propose instead to use cross-entropy loss for DC. Both quantitative and qualitative evaluation are conducted on benchmarks, and we show that switching out sparse depth input and MSE loss functions with our DC representation and loss is a simple way to improve performance, reduce pixel depth mixing and can improve object detection.

To simultaneously estimate head counts and localize heads with bounding boxes, a regression guided detection network (RDNet) is proposed for RGB-D crowd counting. Specifically, to improve the robustness of detection-based approaches for small/tiny heads, we leverage density map to improve the head/non-head classification in detection network where density map serves as the probability of a pixel being a head. A depth-adaptive kernel that considers the variances in head sizes is also introduced to generate high-fidelity density map for more robust density map regression. Further, a depth-aware anchor is designed for better initialization of anchor sizes in detection framework. Then we use the bounding boxes whose sizes are estimated with depth to train our RDNet. The existing RGB-D datasets are too small and not suitable for performance evaluation on data-driven based approaches, we collect a large-scale RGB-D crowd counting dataset. Experiments on both our RGB-D dataset and the MICC RGB-D counting dataset show that our method achieves the best performance for RGB-D crowd counting and localization. Further, our method can be readily extended to RGB image based crowd counting and achieves comparable performance on the ShanghaiTech Part_B dataset for both counting and localization.

We propose a densely semantically aligned person re-identification (re-ID) framework. It fundamentally addresses the body misalignment problem caused by pose/viewpoint variations, imperfect person detection, occlusion, etc.. By leveraging the estimation of the dense semantics of a person image, we construct a set of densely semantically aligned part images (DSAP-images), where the same spatial positions have the same semantics across different person images. We design a two-stream network that consists of a main full image stream (MF-Stream) and a densely semantically-aligned guiding stream (DSAG-Stream). The DSAG-Stream, with the DSAP-images as input, acts as a regulator to guide the MF-Stream to learn densely semantically aligned features from the original image. In the inference, the DSAG-Stream is discarded and only the MF-Stream is needed, which makes the inference system computationally efficient and robust. To our best knowledge, we are the first to make use of fine grained semantics for addressing misalignment problems for re-ID. Our method achieves rank-1 accuracy of 78.9% (new protocol) on the CUHK03 dataset, 90.4% on the CUHK01 dataset, and 95.7% on the Market1501 dataset, outperforming state-of-the-art methods.

A key technical challenge in performing 6D object pose estimation from RGB-D image is to fully leverage the two complementary data sources. Prior works either extract information from the RGB image and depth separately or use costly post-processing steps, limiting their performances in highly cluttered scenes and real-time applications. In this work, we present DenseFusion, a generic framework for estimating 6D pose of a set of known objects from RGB-D images. DenseFusion is a heterogeneous architecture that processes the two data sources individually and uses a novel dense fusion network to extract pixel-wise dense feature embedding, from which the pose is estimated. Furthermore, we integrate an end-to-end iterative pose refinement procedure that further improves the pose estimation while achieving near real-time inference. Our experiments show that our method outperforms state-of-the-art approaches in two datasets, YCB-Video and LineMOD. We also deploy our proposed method to a real robot to grasp and manipulate objects based on the estimated pose.

Our goal in this work is to train an image captioning model that generates more dense and informative captions. We introduce "relational captioning," a novel image captioning task which aims to generate multiple captions with respect to relational information between objects in an image. Relational captioning is a framework that is advantageous in both diversity and amount of information, leading to image understanding based on relationships. Part-of-speech (POS, i.e. subject-object-predicate categories) tags can be assigned to every English word. We leverage the POS as a prior to guide the correct sequence of words in a caption. To this end, we propose a multi-task triple-stream network (MTTSNet) which consists of three recurrent units for the respective POS and jointly performs POS prediction and captioning. We demonstrate more diverse and richer representations generated by the proposed model against several baselines and competing methods.

We present a novel approach for the task of human pose transfer, which aims at synthesizing a new image of a person from an input image of that person and a target pose. We address the issues of limited correspondences identified between keypoints only and invisible pixels due to self-occlusion. Unlike existing methods, we propose to estimate dense and intrinsic 3D appearance flow to better guide the transfer of pixels between poses. In particular, we wish to generate the 3D flow from just the reference and target poses. Training a network for this purpose is non-trivial, especially when the annotations for 3D appearance flow are scarce by nature. We address this problem through a flow synthesis stage. This is achieved by fitting a 3D model to the given pose pair and project them back to the 2D plane to compute the dense appearance flow for training. The synthesized ground-truths are then used to train a feedforward network for efficient mapping from the input and target skeleton poses to the 3D appearance flow. With the appearance flow, we perform feature warping on the input image and generate a photorealistic image of the target pose. Extensive results on DeepFashion and Market-1501 datasets demonstrate the effectiveness of our approach over existing methods. Our code is available at http://mmlab.ie.cuhk.edu.hk/projects/pose-transfer

We present a deep learning system to infer the posterior distribution of a dense depth map associated with an image, by exploiting sparse range measurements, for instance from a lidar. While the lidar may provide a depth value for a small percentage of the pixels, we exploit regularities reflected in the training set to complete the map so as to have a probability over depth for each pixel in the image. We exploit a Conditional Prior Network, that allows associating a probability to each depth value given an image, and combine it with a likelihood term that uses the sparse measurements. Optionally we can also exploit the availability of stereo during training, but in any case only require a single image and a sparse point cloud at run-time. We test our approach on both unsupervised and supervised depth completion using the KITTI benchmark, and improve the state-of-the-art in both.

Few-shot learning for deep neural networks is a highly challenging and key problem in many computer vision tasks. In this context, we are targeting knowledge transfer from a set with abundant data to other sets with few available examples. We propose two simple and effective solutions: (i) dense classification over feature maps, which for the first time studies local activations in the domain of few-shot learning, and (ii) implanting, that is, attaching new neurons to a previously trained network to learn new, task-specific features. Implanting enables training of multiple layers in the few-shot regime, departing from most related methods derived from metric learning that train only the final layer. Both contributions show consistent gains when used individually or jointly and we report state of the art performance on few-shot classification on miniImageNet.