CVPR 2019 Papers

We introduce a novel network, called as CO-attention Siamese Network (COSNet), to address the unsupervised video object segmentation task from a holistic view. We emphasize the importance of inherent correlation among video frames and incorporate a global co-attention mechanism to improve further the state-of-the-art deep learning based solutions that primarily focus on learning discriminative foreground representations over appearance and motion in short-term temporal segments. The co-attention layers in our network provide efficient and competent stages for capturing global correlations and scene context by jointly computing and appending co-attention responses into a joint feature space. We train COSNet with pairs of video frames, which naturally augments training data and allows increased learning capacity. During the segmentation stage, the co-attention model encodes useful information by processing multiple reference frames together, which is leveraged to infer the frequently reappearing and salient foreground objects better. We propose a unified and end-to-end trainable framework where different co-attention variants can be derived for mining the rich context within videos. Our extensive experiments over three large benchmarks manifest that COSNet outperforms the current alternatives by a large margin. We will publicly release our implementation and models.

Recently, deep convolutional neural networks (CNNs) have been widely explored in single image super-resolution (SISR) and obtained remarkable performance. However, most of the existing CNN-based SISR methods mainly focus on wider or deeper architecture design, neglecting to explore the feature correlations of intermediate layers, hence hindering the representational power of CNNs. To address this issue, in this paper, we propose a second-order attention network (SAN) for more powerful feature expression and feature correlation learning. Specifically, a novel train- able second-order channel attention (SOCA) module is developed to adaptively rescale the channel-wise features by using second-order feature statistics for more discriminative representations. Furthermore, we present a non-locally enhanced residual group (NLRG) structure, which not only incorporates non-local operations to capture long-distance spatial contextual information, but also contains repeated local-source residual attention groups (LSRAG) to learn increasingly abstract feature representations. Experimental results demonstrate the superiority of our SAN network over state-of-the-art SISR methods in terms of both quantitative metrics and visual quality.

Conventional neural architecture search (NAS) approaches are usually based on reinforcement learning or evolutionary strategy, which take more than 1000 GPU hours to find a good model on CIFAR-10. We propose an efficient NAS approach, which learns the searching approach by gradient descent. Our approach represents the search space as a directed acyclic graph (DAG). This DAG contains thousands of sub-graphs, each of which indicates a kind of neural architecture. To avoid traversing all the possibilities of the sub-graphs, we develop a differentiable sampler over the DAG. This sampler is learnable and optimized by the validation loss after training the sampled architecture. In this way, our approach can be trained in an end-to-end fashion by gradient descent, named Gradient-based search using Differentiable Architecture Sampler (GDAS). In experiments, we can finish one searching procedure in four GPU hours on CIFAR-10, and the discovered model obtains a test error of 2.82% with only 2.5M parameters, which is on par with the state-of-the-art.

In this work we introduce a novel, CNN-based architecture that can be trained end-to-end to deliver seamless scene segmentation results. Our goal is to predict consistent semantic segmentation and detection results by means of a panoptic output format, going beyond the simple combination of independently trained segmentation and detection models. The proposed architecture takes advantage of a novel segmentation head that seamlessly integrates multi-scale features generated by a Feature Pyramid Network with contextual information conveyed by a light-weight DeepLab-like module. As additional contribution we review the panoptic metric and propose an alternative that overcomes its limitations when evaluating non-instance categories. Our proposed network architecture yields state-of-the-art results on three challenging street-level datasets, i.e. Cityscapes, Indian Driving Dataset and Mapillary Vistas.

Robust recovery of lost colors in underwater images remains a challenging problem. We recently showed that this was partly due to the prevalent use of an atmospheric image formation model for underwater images. We proposed a physically accurate model that explicitly showed: 1) the attenuation coefficient of the signal is not uniform across the scene but depends on object range and reflectance, 2) the coefficient governing the increase in backscatter with distance differs from the signal attenuation coefficient. Here, we present a method that recovers color with the revised model using RGBD images. The Sea-thru method first calculates backscatter using the darkest pixels in the image and their known range information. Then, it uses an estimate of the spatially varying illuminant to obtain the range-dependent attenuation coefficient. Using more than 1,100 images from two optically different water bodies, which we make available, we show that our method outperforms those using the atmospheric model. Consistent removal of water will open up large underwater datasets to powerful computer vision and machine learning algorithms, creating exciting opportunities for the future of underwater exploration and conservation.

Current state-of-the-art object objectors are fine-tuned from the off-the-shelf networks pretrained on large-scale classification dataset ImageNet, which incurs some additional problems: 1) The classification and detection have different degrees of sensitivity to translation, resulting in the learning objective bias; 2) The architecture is limited by the classification network, leading to the inconvenience of modification. To cope with these problems, training detectors from scratch is a feasible solution. However, the detectors trained from scratch generally perform worse than the pretrained ones, even suffer from the convergence issue in training. In this paper, we explore to train object detectors from scratch robustly. By analysing the previous work on optimization landscape, we find that one of the overlooked points in current trained-from-scratch detector is the BatchNorm. Resorting to the stable and predictable gradient brought by BatchNorm, detectors can be trained from scratch stably while keeping the favourable performance independent to the network architecture. Taking this advantage, we are able to explore various types of networks for object detection, without suffering from the poor convergence. By extensive experiments and analyses on downsampling factor, we propose the Root-ResNet backbone network, which makes full use of the information from original images. Our ScratchDet achieves the state-of-the-art accuracy on PASCAL VOC 2007, 2012 and MS COCO among all the train-from-scratch detectors and even performs better than several one-stage pretrained methods. Codes will be made publicly available at https://github.com/KimSoybean/ScratchDet.

Systems which incrementally create 3D semantic maps from image sequences must store and update representations of both geometry and semantic entities. However, while there has been much work on the correct formulation for geometrical estimation, state-of-the-art systems usually rely on simple semantic representations which store and update independent label estimates for each surface element (depth pixels, surfels, or voxels). Spatial correlation is discarded, and fused label maps are incoherent and noisy. We introduce a new compact and optimisable semantic representation by training a variational auto-encoder that is conditioned on a colour image. Using this learned latent space, we can tackle semantic label fusion by jointly optimising the low-dimenional codes associated with each of a set of overlapping images, producing consistent fused label maps which preserve spatial correlation. We also show how this approach can be used within a monocular keyframe based semantic mapping system where a similar code approach is used for geometry. The probabilistic formulation allows a flexible formulation where we can jointly estimate motion, geometry and semantics in a unified optimisation.

Scene Parsing is a challenging task in computer vision, which can be formulated as a pixel-wise classification problem. Existing deep-learning-based methods usually use one general classifier to recognize all object categories. However, the general classifier easily makes some mistakes in dealing with some confusing categories that share similar appearances or semantics. In this paper, we propose an integrated classification model and a variance-based regularization to achieve more accurate classifications. On the one hand, the integrated classification model contains multiple classifiers, not only the general classifier but also a refinement classifier to distinguish the confusing categories. On the other hand, the variance-based regularization differentiates the scores of all categories as large as possible to reduce misclassifications. Specifically, the integrated classification model includes three steps. The first is to extract the features of each pixel. Based on the features, the second step is to classify each pixel across all categories to generate a preliminary classification result. In the third step, we leverage a refinement classifier to refine the classification result, focusing on differentiating the high-preliminary-score categories. An integrated loss with the variance-based regularization is used to train the model. Extensive experiments on three common scene parsing datasets demonstrate the effectiveness of the proposed method.

Many robotic applications require the agent to perform long-horizon tasks in partially observable environments. In such applications, decision making at any step can depend on observations received far in the past. Hence, being able to properly memorize and utilize the long-term history is crucial. In this work, we propose a novel memory-based policy, named Scene Memory Transformer (SMT). The proposed policy embeds and adds each observation to a memory and uses the attention mechanism to exploit spatio-temporal dependencies. This model is generic and can be efficiently trained with reinforcement learning over long episodes. On a range of visual navigation tasks, SMT demonstrates superior performance to existing reactive and memory-based policies by a margin.

Scene graph generation has received growing attention with the advancements in image understanding tasks such as object detection, attributes and relationship prediction, etc. However, existing datasets are biased in terms of object and relationship labels, or often come with noisy and missing annotations, which makes the development of a reliable scene graph prediction model very challenging. In this paper, we propose a novel scene graph generation algorithm with external knowledge and image reconstruction loss to overcome these dataset issues. In particular, we extract commonsense knowledge from the external knowledge base to refine object and phrase features for improving generalizability in scene graph generation. To address the bias of noisy object annotations, we introduce an auxiliary image reconstruction path to regularize the scene graph generation network. Extensive experiments show that our framework can generate better scene graphs, achieving the state-of-the-art performance on two benchmark datasets: Visual Relationship Detection and Visual Genome datasets.

The computer vision community has witnessed recent advances in scene categorization from images, with the state of the art systems now achieving impressive recognition rates on challenging benchmarks. Such systems have been trained on photographs which include color, texture and shading cues. The geometry of shapes and surfaces, as conveyed by scene contours, is not explicitly considered for this task. Remarkably, humans can accurately recognize natural scenes from line drawings, which consist solely of contour-based shape cues. Here we report the first computer vision study on scene categorization of line drawings derived from popular databases including an artist scene database, MIT67 and Places365. Specifically, we use off-the-shelf pre-trained Convolutional Neural Networks (CNNs) to perform scene classification given only contour information as input, and find performance levels well above chance. We also show that medial-axis based contour salience methods can be used to select more informative subsets of contour pixels, and that the variation in CNN classification performance on various choices for these subsets is qualitatively similar to that observed in human performance. Moreover, when the salience measures are used to weight the contours, we find that these weights boost our CNN performance above that for unweighted contour input. That is, the medial axis based salience weights appear to add useful information that is not available when CNNs are trained to use contours alone.

We introduce Scan2Mesh, a novel data-driven generative approach which transforms an unstructured and potentially incomplete range scan into a structured 3D mesh representation. The main contribution of this work is a generative neural network architecture whose input is a range scan of a 3D object and whose output is an indexed face set conditioned on the input scan. In order to generate a 3D mesh as a set of vertices and face indices, the generative model builds on a series of proxy losses for vertices, edges, and faces. At each stage, we realize a one-to-one discrete mapping between the predicted and ground truth data points with a combination of convolutional- and graph neural network architectures. This enables our algorithm to predict a compact mesh representation similar to those created through manual artist effort using 3D modeling software. Our generated mesh results thus produce sharper, cleaner meshes with a fundamentally different structure from those generated through implicit functions, a first step in bridging the gap towards artist-created CAD models.

We present Scan2CAD, a novel data-driven method that learns to align clean 3D CAD models from a shape database to the noisy and incomplete geometry of a commodity RGB-D scan. For a 3D reconstruction of an indoor scene, our method takes as input a set of CAD models, and predicts a 9DoF pose that aligns each model to the underlying scan geometry. To tackle this problem, we create a new scan-to-CAD alignment dataset based on 1506 ScanNet scans with 97607 annotated keypoint pairs between 14225 CAD models from ShapeNet and their counterpart objects in the scans. Our method selects a set of representative keypoints in a 3D scan for which we find correspondences to the CAD geometry. To this end, we design a novel 3D CNN architecture that learns a joint embedding between real and synthetic objects, and from this predicts a correspondence heatmap. Based on these correspondence heatmaps, we formulate a variational energy minimization that aligns a given set of CAD models to the reconstruction. We evaluate our approach on our newly introduced Scan2CAD benchmark where we outperform both handcrafted feature descriptor as well as state-of-the-art CNN based methods by 21.39%.

How do computers and intelligent agents view the world around them? Feature extraction and representation constitutes one the basic building blocks towards answering this question. Traditionally, this has been done with carefully engineered hand-crafted techniques such as HOG, SIFT or ORB. However, there is no "one size fits all" approach that satisfies all requirements. In recent years, the rising popularity of deep learning has resulted in a myriad of end-to-end solutions to many computer vision problems. These approaches, while successful, tend to lack scalability and can't easily exploit information learned by other systems. Instead, we propose SAND features, a dedicated deep learning solution to feature extraction capable of providing hierarchical context information. This is achieved by employing sparse relative labels indicating relationships or similarity of dissimilarity between image locations. The nature of these labels results in an almost infinite set of dissimilar examples to choose from. We demonstrate how the selection of negative examples during training can be used to modify the feature space and vary it's learned properties. To demonstrate the generality of this approach, we apply the proposed features to a multitude of tasks, each requiring different properties. This includes disparity estimation, semantic segmentation, self-localisation and SLAM. In all cases, we show how incorporating SAND features results in better or comparable results to the baseline, whilst requiring little to no additional training.

Recently, deep learning based image Compressed Sensing (CS) methods have been proposed and demonstrated superior reconstruction quality with low computational complexity. However, the existing deep learning based image CS methods need to train different models for different sampling ratios, which increases the complexity of the encoder and decoder. In this paper, we propose a scalable convolutional neural network (dubbed SCSNet) to achieve scalable sampling and scalable reconstruction with only one model. Specifically, SCSNet provides both coarse and fine granular scalability. For coarse granular scalability, SCSNet is designed as a single sampling matrix plus a hierarchical reconstruction network that contains a base layer plus multiple enhancement layers. The base layer provides the basic reconstruction quality, while the enhancement layers reference the lower reconstruction layers and gradually improve the reconstruction quality. For fine granular scalability, SCSNet achieves sampling and reconstruction at any sampling ratio by using a greedy method to select the measurement bases. Compared with the existing deep learning based image CS methods, SCSNet achieves scalable sampling and quality scalable reconstruction at any sampling ratio with only one model. Experimental results demonstrate that SCSNet has the state-of-the-art performance while maintaining a comparable running speed with the existing deep learning based image CS methods.

Efficient and reliable methods for training of object detectors are in higher demand than ever, and more and more data relevant to the field is becoming available. However, large datasets like Open Images Dataset v4 (OID) are sparsely annotated, and some measure must be taken in order to ensure the training of a reliable detector. In order to take the incompleteness of these datasets into account, one possibility is to use pretrained models to detect the presence of the unverified objects. However, the performance of such a strategy depends largely on the power of the pretrained model. In this study, we propose part-aware sampling, a method that uses human intuition for the hierarchical relation between objects. In terse terms, our method works by making assumptions like "a bounding box for a car should contain a bounding box for a tire". We demonstrate the power of our method on OID and compare the performance against a method based on a pretrained model. Our method also won the first and second place on the public and private test sets of the Google AI Open Images Competition 2018.

This paper presents a new method for detecting salient objects in images using convolutional neural networks (CNNs). The proposed network, named PAGE-Net, offers two key contributions. The first is the exploitation of an essential pyramid attention structure for salient object detection. This enables the network to concentrate more on salient regions while considering multi-scale saliency information. Such a stacked attention design provides a powerful tool to efficiently improve the representation ability of the corresponding network layer with an enlarged receptive field. The second contribution lies in the emphasis on the importance of salient edges. Salient edge information offers a strong cue to better segment salient objects and refine object boundaries. To this end, our model is equipped with a salient edge detection module, which is learned for precise salient boundary estimation. This encourages better edge-preserving salient object segmentation. Exhaustive experiments confirm that the proposed pyramid attention and salient edges are effective for salient object detection. We show that our deep saliency model outperforms state-of-the-art approaches for several benchmarks with a fast processing speed (25fps on one GPU).

Arbitrary attribute editing generally can be tackled by incorporating encoder-decoder and generative adversarial networks. However, the bottleneck layer in encoder-decoder usually gives rise to blurry and low quality editing result. And adding skip connections improves image quality at the cost of weakened attribute manipulation ability. Moreover, existing methods exploit target attribute vector to guide the flexible translation to desired target domain. In this work, we suggest to address these issues from selective transfer perspective. Considering that specific editing task is certainly only related to the changed attributes instead of all target attributes, our model selectively takes the difference between target and source attribute vectors as input. Furthermore, selective transfer units are incorporated with encoder-decoder to adaptively select and modify encoder feature for enhanced attribute editing. Experiments show that our method (i.e., STGAN) simultaneously improves attribute manipulation accuracy as well as perception quality, and performs favorably against state-of-the-arts in arbitrary face attribute editing and season translation.

In this paper, we propose Spatio-TEmporal Progressive (STEP) action detector--a progressive learning framework for spatio-temporal action detection in videos. Starting from a handful of coarse-scale proposal cuboids, our approach progressively refines the proposals towards actions over a few steps. In this way, high-quality proposals (i.e., adhere to action movements) can be gradually obtained at later steps by leveraging the regression outputs from previous steps. At each step, we adaptively extend the proposals in time to incorporate more related temporal context. Compared to the prior work that performs action detection in one run, our progressive learning framework is able to naturally handle the spatial displacement within action tubes and therefore provides a more effective way for spatio-temporal modeling. We extensively evaluate our approach on UCF101 and AVA, and demonstrate superior detection results. Remarkably, we achieve mAP of 75.0% and 18.6% on the two datasets with 3 progressive steps and using respectively only 11 and 34 initial proposals.