WACV 2021 论文列表

We propose a self-supervised approach for training multi-frame video denoising networks. These networks predict frame t from a window of frames around t. Our self-supervised approach benefits from the video temporal consistency by penalizing a loss between the predicted frame t and a neighboring target frame, which are aligned using an optical flow. We use the proposed strategy for online internal learning, where a pre-trained network is fine-tuned to denoise a new unknown noise type from a single video. After a few frames, the proposed fine-tuning reaches and sometimes surpasses the performance of a state-of-the-art network trained with supervision. In addition, for a wide range of noise types, it can be applied blindly without knowing the noise distribution. We demonstrate this by showing results on blind denoising of different synthetic and real noises.

We extend the blindspot model for self-supervised denoising to handle Poisson-Gaussian noise and introduce an improved training scheme that avoids hyperparameters and adapts the denoiser to the test data. Self-supervised models for denoising learn to denoise from only noisy data and do not require corresponding clean images, which are difficult or impossible to acquire in some application areas of interest such as low-light microscopy. We introduce a new training strategy to handle Poisson-Gaussian noise which is the standard noise model for microscope images. Our new strategy eliminates hyperparameters from the loss function, which is important in a self-supervised regime where no ground truth data is available to guide hyperparameter tuning. We show how our denoiser can be adapted to the test data to improve performance. Our evaluations on microscope image denoising benchmarks validate our approach.

Self-supervised learning (SSL) is a technique for learning useful representations from unlabeled data. It has been applied effectively to domain adaptation (DA) on images and videos. It is still unknown if and how it can be leveraged for domain adaptation in 3D perception problems. Here we describe the first study of SSL for DA on point clouds. We introduce a new family of pretext tasks, Deformation Reconstruction, inspired by the deformations encountered in sim-to-real transformations. In addition, we propose a novel training procedure for labeled point cloud data motivated by the MixUp method called Point cloud Mixup (PCM). Evaluations on domain adaptations datasets for classification and segmentation, demonstrate a large improvement over existing and baseline methods.

Recently 3D scene understanding attracts attention for many applications, however, annotating a vast amount of 3D data for training is usually expensive and time-consuming. To alleviate the needs of ground truth, we propose a self-supervised schema to learn 4D spatio-temporal features (i.e. 3 spatial dimensions plus 1 temporal dimension) from dynamic point cloud data by predicting the temporal order of sampled and shuffled point cloud clips. 3D sequential point cloud contains precious geometric and depth information to better recognize activities in 3D space compared to videos. To learn the 4D spatio-temporal features, we introduce 4D convolution neural networks to predict the temporal order on a self-created large scale dataset, NTU-PCLs, derived from the NTU-RGB+D dataset. The efficacy of the learned 4D spatio-temporal features is verified on two tasks: 1) Self-supervised 3D nearest neighbor retrieval; and 2) Self-supervised representation learning transferred for action recognition on the smaller 3D dataset. Our extensive experiments prove the effectiveness of the proposed self-supervised learning method which achieves comparable results w.r.t. the fully-supervised methods on action recognition on MSRAction3D dataset.

The development of large-scale image-captioning datasets is expensive, while the abundance of unpaired images and text corpus can potentially help reduce the efforts of manual annotation. In this paper, we study the few-shot image captioning problem that only requires a small amount of annotated image-caption pairs. We propose an ensemble-based self-distillation method that allows image captioning models to be trained with unpaired images and captions. The ensemble consists of multiple base models trained with different data samples in each iteration. For learning from unpaired images, we generate multiple pseudo captions with the ensemble and allocate different weights according to their confidence levels. For learning from unpaired captions, we propose a simple yet effective pseudo feature generation method based on Gradient Descent. The pseudo captions and pseudo features from the ensemble are used to train the base models in future iterations. The proposed method is general over different image captioning models and datasets. Our experiments demonstrate significant performance improvements and meaningful captions generated with only 1% of paired training data. Source code is available at https://github.com/chenxy99/SD-FSIC.

In recent years, the attention mechanism has become a fairly popular concept and has proven to be successful in many machine learning applications. However, deep learning models do not employ supervision for these attention mechanisms which can improve the model's performance significantly. Therefore, in this paper, we tackle this limitation and propose a novel method to improve the attention mechanism by inducing "self-supervision". We devise a technique to generate desirable attention maps for any model that utilizes an attention module. This is achieved by examining the model's output for different regions sampled from the input and obtaining the attention probability distributions that enhance the proficiency of the model. The attention distributions thus obtained are used for supervision. We rely on the fact, that attenuation of the unimportant parts, allows a model to attend to more salient regions, thus strengthening the prediction accuracy. The quantitative and qualitative results published in this paper show that this method successfully improves the attention mechanism as well as the model's accuracy. In addition to the task of Visual Question Answering(VQA), we also show results on the task of Image classification and Text classification to prove that our method can be generalized to any vision and language model that uses an attention module

Taking advantage of human pose data for understanding human activities has attracted much attention these days. However, state-of-the-art pose estimators struggle in obtaining high-quality 2D or 3D pose data due to occlusion, truncation and low-resolution in real-world un-annotated videos. Hence, in this work, we propose 1) a Selective Spatio-Temporal Aggregation mechanism, named SST-A, that refines and smooths the keypoint locations extracted by multiple expert pose estimators, 2) an effective weakly-supervised self-training framework which leverages the aggregated poses as pseudo ground-truth instead of handcrafted annotations for real-world pose estimation. Extensive experiments are conducted for evaluating not only the upstream pose refinement but also the downstream action recognition performance on four datasets, Toyota Smarthome, NTU-RGB+D, Charades, and Kinetics-50. We demonstrate that the skeleton data refined by our Pose-Refinement system (SSTA-PRS) is effective at boosting various existing action recognition models, which achieves competitive or state-of-the-art performance. Refined pose data is available at: https://github.com/walker-a11y/SSTA-PRS

We introduce a new class of vision-based sensor and associated algorithmic processes that combine visual imaging with high-resolution tactile sending, all in a uniform hardware and computational architecture. We demonstrate the sensor's efficacy for both multi-modal object recognition and metrology. Object recognition is typically formulated as an unimodal task, but by combining two sensor modalities we show that we can achieve several significant performance improvements. This sensor, named the See-Through-your-Skin sensor (STS), is designed to provide rich multi-modal sensing of contact surfaces. Inspired by recent developments in optical tactile sensing technology, we address a key missing feature of these sensors: the ability to capture a visual perspective of the region beyond the contact surface. Whereas optical tactile sensors are typically opaque, we present a sensor with a semitransparent skin that has the dual capabilities of acting as a tactile sensor and/or as a visual camera depending on its internal lighting conditions. This paper details the design of the sensor, showcases its dual sensing capabilities, and presents a deep learning architecture that fuses vision and touch. We validate the ability of the sensor to classify household objects, recognize fine textures, and infer their physical properties both through numerical simulations and experimentally with a smart countertop prototype.

Using only a mobile phone app, our objective is to cheaply retro-fit digital meters (e.g blood pressure, blood glucose or industrial gauges) with `smart' data transfer capabilities. Using the mobile phone camera we build an app to securely and accurately transcribe information from digital meter screens. Only a single labelled training image of a target meter is required to build a custom screen reading module. Here we show how this can scale to potentially hundreds of different meters by learning to recognising the meter type so that the reading module can be automatically selected. This makes the system very easy for a user who would need to scan multiple different meter types. To this end, we build a CNN based system which runs in real-time on mobile device with very high read accuracy and meter recognition. Our contributions include (i) a method of one-shot training by synthesis through domain shift reduction, (ii) a deep embedding network for scale, translation and rotation invariant re-identification of digital meters, (iii) a highly accurate and efficient mobile phone app for recognising and parsing digital meter screens and (iv) release of a new digital meter re-identification dataset.

Siamese trackers turn tracking into similarity estimation between a template and the candidate regions in the frame. Mathematically, one of the key ingredients of success of the similarity function is translation equivariance. Non-translation-equivariant architectures induce a positional bias during training, so the location of the target will be hard to recover from the feature space. In real life scenarios, objects undergo various transformations other than translation, such as rotation or scaling. Unless the model has an internal mechanism to handle them, the similarity may degrade. In this paper, we focus on scaling and we aim to equip the Siamese network with additional built-in scale equivariance to capture the natural variations of the target a priori. We develop the theory for scale-equivariant Siamese trackers, and provide a simple recipe for how to make a wide range of existing trackers scale-equivariant. We present SE-SiamFC, a scale-equivariant variant of SiamFC built according to the recipe. We conduct experiments on OTB and VOT benchmarks and on the synthetically generated T-MNIST and S-MNIST datasets. We demonstrate that a built-in additional scale equivariance is useful for visual object tracking.

Land-cover classification using remote sensing imagery is an important Earth observation task. Recently, land cover classification has benefited from the development of fully connected neural networks for semantic segmentation. The benchmark datasets available for training deep segmentation models in remote sensing imagery tend to be small, however, often consisting of only a handful of images from a single location with a single scale. This limits the models' ability to generalize to other datasets. Domain adaptation has been proposed to improve the models' generalization but we find these approaches are not effective for dealing with the scale variation commonly found between remote sensing image collections. We therefore propose a scale aware adversarial learning framework to perform joint cross-location and cross-scale land-cover classification. The framework has a dual discriminator architecture with a standard feature discriminator as well as a novel scale discriminator. We also introduce a scale attention module which produces scale-enhanced features. Experimental results show that the proposed framework outperforms state-of-the-art domain adaptation methods by a large margin. The open-sourced codes are available on Github: https://github.com/xdeng7/scale-aware_da.

The detection of manufacturing errors is crucial in fabrication processes to ensure product quality and safety standards. Since many defects occur very rarely and their characteristics are mostly unknown a priori, their detection is still an open research question. To this end, we propose DifferNet: It leverages the descriptiveness of features extracted by convolutional neural networks to estimate their density using normalizing flows. Normalizing flows are well-suited to deal with low dimensional data distributions. However, they struggle with the high dimensionality of images. Therefore, we employ a multi-scale feature extractor which enables the normalizing flow to assign meaningful likelihoods to the images. Based on these likelihoods we develop a scoring function that indicates defects. Moreover, propagating the score back to the image enables pixel-wise localization. To achieve a high robustness and performance we exploit multiple transformations in training and evaluation. In contrast to most other methods, ours does not require a large number of training samples and performs well with as low as 16 images. We demonstrate the superior performance over existing approaches on the challenging and newly proposed MVTec AD and Magnetic Tile Defects datasets.

The last decades have seen great progress in saliency prediction, with the success of deep neural networks that are able to encode high-level semantics. Yet, while humans have the innate capability in leveraging their knowledge to decide where to look (e.g. people pay more attention to familiar faces such as celebrities), saliency prediction models have only been trained with large eye-tracking datasets. This work proposes to bridge this gap by explicitly incorporating external knowledge for saliency models as humans do. We develop networks that learn to highlight regions by incorporating prior knowledge of semantic relationships, be it general or domain-specific, depending on the task of interest. At the core of the method is a new GraSSNet that constructs a graph that encodes semantic relationships learned from external knowledge. A Spatial Graph Attention Network is then developed to update saliency features based on the learned graph. Experiments show that the proposed model learns to predict saliency from the external knowledge and outperforms the state-of-the-art on four saliency benchmarks.

This paper proposes a new end-to-end trainable model for lossy image compression, which includes several novel components. The method incorporates 1) an adequate perceptual similarity metric; 2) saliency in the images; 3) a hierarchical auto-regressive model. This paper demonstrates that the popularly used evaluations metrics such as MS-SSIM and PSNR are inadequate for judging the performance of image compression techniques as they do not align with the human perception of similarity. Alternatively, a new metric is proposed, which is learned on perceptual similarity data specific to image compression. The proposed compression model incorporates the salient regions and optimizes on the proposed perceptual similarity metric. The model not only generates images which are visually better but also gives superior performance for subsequent computer vision tasks such as object detection and segmentation when compared to existing engineered or learned compression techniques.

Providing an explanation of the operation of CNNs that is both accurate and interpretable is becoming essential in fields like medical image analysis, surveillance, and autonomous driving. In these areas, it is important to have confidence that the CNN is working as expected and explanations from saliency maps provide an efficient way of doing this. In this paper, we propose a pair of complementary contributions that improve upon the state of the art for region-based explanations in both accuracy and utility. The first is SWAG, a method for generating accurate explanations quickly using superpixels for discriminative regions which is meant to be a more accurate, efficient, and tunable drop in replacement method for Grad-CAM, LIME, or other region-based methods. The second contribution is based on an investigation into how to best generate the superpixels used to represent the features found within the image. Using SWAG, we compare using superpixels created from the image, a combination of the image and backpropagated gradients, and the gradients themselves. To the best of our knowledge, this is the first method proposed to generate explanations using superpixels explicitly created to represent the discriminative features important to the network. To compare we use both ImageNet and challenging fine-grained datasets over a range of metrics. We demonstrate experimentally that our methods provide the best local and global accuracy compared to Grad-CAM, Grad-CAM++, LIME, XRAI, and RISE.

Interpolation of sparse pixel information towards a dense target resolution finds its application across multiple disciplines in computer vision. State-of-the-art interpolation of motion fields applies model-based interpolation that makes use of edge information extracted from the target image. For depth completion, data-driven learning approaches are widespread. Our work is inspired by latest trends in depth completion that tackle the problem of dense guidance for sparse information. We extend these ideas and create a generic cross-domain architecture that can be applied for a multitude of interpolation problems like optical flow, scene flow, or depth completion. In our experiments, we show that our proposed concept of Sparse Spatial Guided Propagation (SSGP) achieves improvements to robustness, accuracy, or speed compared to specialized algorithms.

Recent advances in deep generative models have led to an unprecedented level of realism for synthetically generated images of humans. However, one of the remaining fundamental limitations of these models is the ability to flexibly control the generative process, e.g. change the camera and human pose while retaining the subject identity. At the same time, deformable human body models like SMPL and its successors provide full control over pose and shape but rely on classic computer graphics pipelines for rendering. Such rendering pipelines require explicit mesh rasterization that (a) does not have the potential to fix artifacts or lack of realism in the original 3D geometry and (b) until recently, were not fully incorporated into deep learning frameworks. In this work, we propose to bridge the gap between classic geometry-based rendering and the latest generative networks operating in pixel space. We train a network that directly converts a sparse set of 3D mesh vertices into photorealistic images, alleviating the need for traditional rasterization mechanism. We train our model on a large corpus of human 3D models and corresponding real photos, and show the advantage over conventional differentiable renderers both in terms of the level of photorealism and rendering efficiency.

This paper demonstrates a system capable of combining a sparse, indirect, monocular visual SLAM, with both offline and real-time Multi-View Stereo (MVS) reconstruction algorithms. This combination overcomes many obstacles encountered by autonomous vehicles or robots employed in agricultural environments, such as overly repetitive patterns, need for very detailed reconstructions, and abrupt movements caused by uneven roads. Furthermore, the use of a monocular SLAM makes our system much easier to integrate with an existing device, as we do not rely on a LiDAR (which is expensive and power consuming), or stereo camera (whose calibration is sensitive to external perturbation e.g. camera being displaced). To the best of our knowledge, this paper presents the first evaluation results for monocular SLAM, and our work further explores unsupervised depth estimation on this specific application scenario by simulating RGB-D SLAM to tackle the scale ambiguity, and shows our approach produces reconstructions that are helpful to various agricultural tasks. Moreover, we highlight that our experiments provide meaningful insight to improve monocular SLAM systems under agricultural settings.

Hatching is a common method used by artists to accentuate the third dimension of a sketch, and to illuminate the scene. Our system SHAD3S attempts to compete with a human at hatching generic three-dimensional (3D) shapes, and also tries to assist her in a form exploration exercise. The novelty of our approach lies in the fact that we make no assumptions about the input other than that it represents a 3D shape, and yet, given a contextual information of illumination and texture, we synthesise an accurate hatch pattern over the sketch, without access to 3d or pseudo 3D. In the process, we contribute towards a) a cheap yet effective method to synthesise a sufficiently large high fidelity dataset, pertinent to task; b) creating a pipeline with conditional generative adversarial network (CGAN); and c) creating an interactive utility with GIMP, that is a tool for artists to engage with automated hatching or a form-exploration exercise. User evaluation of the tool suggests that the model performance does generalise satisfactorily over diverse input, both in terms of style as well as shape. A simple comparison of inception scores suggest that the generated distribution is as diverse as the ground truth.k