Current benchmarks for optical flow algorithms evaluate the estimation quality by comparing their predicted flow field with the ground truth, and additionally may compare interpolated frames, based on these predictions, with the correct frames from the actual image sequences. For the latter comparisons, objective measures such as mean square errors are applied. However, for applications like image interpolation, the expected user's quality of experience cannot be fully deduced from such simple quality measures. Therefore, we conducted a subjective quality assessment study by crowdsourcing for the interpolated images provided in one of the optical flow benchmarks, the Middlebury benchmark. We used paired comparisons with forced choice and reconstructed absolute quality scale values according to Thurstone's model using the classical least squares method. The results give rise to a re-ranking of 141 participating algorithms w.r.t. visual quality of interpolated frames mostly based on optical flow estimation. Our re-ranking result shows the necessity of visual quality assessment as another evaluation metric for optical flow and frame interpolation benchmarks.
While being the de facto standard coordinate representation in human pose estimation, heatmap is never systematically investigated in the literature, to our best knowledge. This work fills this gap by studying the coordinate representation with a particular focus on the heatmap. Interestingly, we found that the process of decoding the predicted heatmaps into the final joint coordinates in the original image space is surprisingly significant for human pose estimation performance, which nevertheless was not recognised before. In light of the discovered importance, we further probe the design limitations of the standard coordinate decoding method widely used by existing methods, and propose a more principled distribution-aware decoding method. Meanwhile, we improve the standard coordinate encoding process (i.e. transforming ground-truth coordinates to heatmaps) by generating accurate heatmap distributions for unbiased model training. Taking the two together, we formulate a novel Distribution-Aware coordinate Representation of Keypoint (DARK) method. Serving as a model-agnostic plug-in, DARK significantly improves the performance of a variety of state-of-the-art human pose estimation models. Extensive experiments show that DARK yields the best results on two common benchmarks, MPII and COCO, consistently validating the usefulness and effectiveness of our novel coordinate representation idea.
We present an unsupervised approach for learning to estimate three dimensional (3D) facial structure from a single image while also predicting 3D viewpoint transformations that match a desired pose and facial geometry. We achieve this by inferring the depth of facial keypoints of an input image in an unsupervised manner, without using any form of ground-truth depth information. We show how it is possible to use these depths as intermediate computations within a new backpropable loss to predict the parameters of a 3D affine transformation matrix that maps inferred 3D keypoints of an input face to the corresponding 2D keypoints on a desired target facial geometry or pose. Our resulting approach can therefore be used to infer plausible 3D transformations from one face pose to another, allowing faces to be frontalized, transformed into 3D models or even warped to another pose and facial geometry. Lastly, we identify certain shortcomings with our formulation, and explore adversarial image translation techniques as a post-processing step to re-synthesize complete head shots for faces re-targeted to different poses or identities.
Learning an effective attention mechanism for multimodal data is important in many vision-and-language tasks that require a synergic understanding of both the visual and textual contents. Existing state-of-the-art approaches use co-attention models to associate each visual object (e.g., image region) with each textual object (e.g., query word). Despite the success of these co-attention models, they only model inter-modal interactions while neglecting intra-modal interactions. Here we propose a general `unified attention' model that simultaneously captures the intra- and inter-modal interactions of multimodal features and outputs their corresponding attended representations. By stacking such unified attention blocks in depth, we obtain the deep Multimodal Unified Attention Network (MUAN), which can seamlessly be applied to the visual question answering (VQA) and visual grounding tasks. We evaluate our MUAN models on two VQA datasets and three visual grounding datasets, and the results show that MUAN achieves top-level performance on both tasks without bells and whistles.
Over the past four years, neural networks have been proven vulnerable to adversarial images: targeted but imperceptible image perturbations lead to drastically different predictions. We show that adversarial vulnerability increases with the gradients of the training objective when viewed as a function of the inputs. For most current network architectures, we prove that the $\ell_1$-norm of these gradients grows as the square root of the input size. These nets therefore become increasingly vulnerable with growing image size. Our proofs rely on the network's weight distribution at initialization, but extensive experiments confirm that our conclusions still hold after training.
Conventional approaches to image registration consist of time consuming iterative methods. Most current deep learning (DL) based registration methods extract deep features to use in an iterative setting. We propose an end-to-end DL method for registering multimodal images. Our approach uses generative adversarial networks (GANs) that eliminates the need for time consuming iterative methods, and directly generates the registered image with the deformation field. Appropriate constraints in the GAN cost function produce accurately registered images in less than a second. Experiments demonstrate their accuracy for multimodal retinal and cardiac MR image registration.
Matrix completion problems are the problems of recovering missing entries in a partially observed high dimensional matrix with or without noise. Such a problem is encountered in a wide range of applications such as collaborative filtering, global positioning and remote sensing. Most of the existing matrix completion algorithms assume a low rank structure of the underlying complete matrix and perform reconstruction through the recovery of the low-rank structure using singular value decomposition. In this paper, we propose an alternative and more flexible structure for the underlying true complete matrix for the purpose of matrix completion and denoising. Specifically, instead of assuming a low matrix rank, we assume the underlying complete matrix has a low Kronecker product rank structure. Such a structure is often seen in the matrix observations in signal processing and image processing applications. The Kronecker product structure also includes low rank singular value decomposition structure commonly used as one of its special cases. The extra flexibility assumed for the underlying structure allows for using much less number of parameters but also raises the challenge of determining the proper Kronecker product configuration to be used. In this article, we propose to use a class of information criteria for the determination of the proper configuration and study its empirical performance in matrix completion problems. Simulation studies show promising results that the true underlying configuration can be accurately selected by the information criteria and the accompanying matrix completion algorithm can produce more accurate matrix recovery with less number of parameters than the standard matrix completion algorithms.
In this work we present a method to train a plane-aware convolutional neural network for dense depth and surface normal estimation as well as plane boundaries from a single indoor \threesixty image. Using our proposed loss function, our network outperforms existing methods for single-view, indoor, omnidirectional depth estimation and provides an initial benchmark for surface normal prediction from \threesixty images. Our improvements are due to the use of a novel plane-aware loss that leverages principal curvature as an indicator of planar boundaries. We also show that including geodesic coordinate maps as network priors provides a significant boost in surface normal prediction accuracy. Finally, we demonstrate how we can combine our network's outputs to generate high quality 3D ``pop-up" models of indoor scenes.
We have successfully implemented the "Learn to Pay Attention" model of attention mechanism in convolutional neural networks, and have replicated the results of the original paper in the categories of image classification and fine-grained recognition.
Depth Completion deals with the problem of converting a sparse depth map to a dense one, given the corresponding color image. Convolutional spatial propagation network (CSPN) is one of the state-of-the-art (SoTA) methods of depth completion, which recovers structural details of the scene. In this paper, we propose CSPN++, which further improves its effectiveness and efficiency by learning adaptive convolutional kernel sizes and the number of iterations for the propagation, thus the context and computational resources needed at each pixel could be dynamically assigned upon requests. Specifically, we formulate the learning of the two hyper-parameters as an architecture selection problem where various configurations of kernel sizes and numbers of iterations are first defined, and then a set of soft weighting parameters are trained to either properly assemble or select from the pre-defined configurations at each pixel. In our experiments, we find weighted assembling can lead to significant accuracy improvements, which we referred to as "context-aware CSPN", while weighted selection, "resource-aware CSPN" can reduce the computational resource significantly with similar or better accuracy. Besides, the resource needed for CSPN++ can be adjusted w.r.t. the computational budget automatically. Finally, to avoid the side effects of noise or inaccurate sparse depths, we embed a gated network inside CSPN++, which further improves the performance. We demonstrate the effectiveness of CSPN++on the KITTI depth completion benchmark, where it significantly improves over CSPN and other SoTA methods.