Optical flow estimation aims to find the 2D motion field by identifying corresponding pixels between two images. Despite the tremendous progress of deep learning-based optical flow methods, it remains a challenge to accurately estimate large displacements with motion blur. This is mainly because the correlation volume, the basis of pixel matching, is computed as the dot product of the convolutional features of the two images. The locality of convolutional features makes the computed correlations susceptible to various noises. On large displacements with motion blur, noisy correlations could cause severe errors in the estimated flow. To overcome this challenge, we propose a new architecture “CRoss-Attentional Flow Trans-former” (CRAFT), aiming to revitalize the correlation volume computation. In CRAFT, a Semantic Smoothing Trans-former layer transforms the features of one frame, making them more global and semantically stable. In addition, the dot-product correlations are replaced with trans-former Cross-Frame Attention. This layer filters out feature noises through the Query and Key projections, and computes more accurate correlations. On Sintel (Final) and KITTI (foreground) benchmarks, CRAFT has achieved new state-of-the-art performance. Moreover, to test the robust-ness of different models on large motions, we designed an image shifting attack that shifts input images to generate large artificial motions. Under this attack, CRAFT per-forms much more robustly than two representative meth-ods, RAFT and GMA. The code of CRAFT is is available at https://github.com/askerlee/craft.

Medical image segmentation is important for computer-aided diagnosis. Good segmentation demands the model to see the big picture and fine details simultaneously, i.e., to learn image features that incorporate large context while keep high spatial resolutions. To approach this goal, the most widely used methods -- U-Net and variants, extract and fuse multi-scale features. However, the fused features still have small "effective receptive fields" with a focus on local image cues, limiting their performance. In this work, we propose Segtran, an alternative segmentation framework based on transformers, which have unlimited "effective receptive fields" even at high feature resolutions. The core of Segtran is a novel Squeeze-and-Expansion transformer: a squeezed attention block regularizes the self attention of transformers, and an expansion block learns diversified representations. Additionally, we propose a new positional encoding scheme for transformers, imposing a continuity inductive bias for images. Experiments were performed on 2D and 3D medical image segmentation tasks: optic disc/cup segmentation in fundus images (REFUGE'20 challenge), polyp segmentation in colonoscopy images, and brain tumor segmentation in MRI scans (BraTS'19 challenge). Compared with representative existing methods, Segtran consistently achieved the highest segmentation accuracy, and exhibited good cross-domain generalization capabilities.

Robotic drawing has become increasingly popular as an entertainment and interactive tool. In this paper we present RoboCoDraw, a real-time collaborative robot-based drawing system that draws stylized human face sketches interactively in front of human users, by using the Generative Adversarial Network (GAN)-based style transfer and a Random-Key Genetic Algorithm (RKGA)-based path optimization. The proposed RoboCoDraw system takes a real human face image as input, converts it to a stylized avatar, then draws it with a robotic arm. A core component in this system is the AvatarGAN proposed by us, which generates a cartoon avatar face image from a real human face. AvatarGAN is trained with unpaired face and avatar images only and can generate avatar images of much better likeness with human face images in comparison with the vanilla CycleGAN. After the avatar image is generated, it is fed to a line extraction algorithm and converted to sketches. An RKGA-based path optimization algorithm is applied to find a time-efficient robotic drawing path to be executed by the robotic arm. We demonstrate the capability of RoboCoDraw on various face images using a lightweight, safe collaborative robot UR5.

Neural Style Transfer based on Convolutional Neural Networks (CNN) aims to synthesize a new image that retains the high-level structure of a content image, rendered in the low-level texture of a style image. This is achieved by constraining the new image to have high-level CNN features similar to the content image, and lower-level CNN features similar to the style image. However in the traditional optimization objective, low-level features of the content image are absent, and the low-level features of the style image dominate the low-level detail structures of the new image. Hence in the synthesized image, many details of the content image are lost, and a lot of inconsistent and unpleasing artifacts appear. As a remedy, we propose to steer image synthesis with a novel loss function: the Laplacian loss. The Laplacian matrix ("Laplacian" in short), produced by a Laplacian operator, is widely used in computer vision to detect edges and contours. The Laplacian loss measures the difference of the Laplacians, and correspondingly the difference of the detail structures, between the content image and a new image. It is flexible and compatible with the traditional style transfer constraints. By incorporating the Laplacian loss, we obtain a new optimization objective for neural style transfer named Lapstyle. Minimizing this objective will produce a stylized image that better preserves the detail structures of the content image and eliminates the artifacts. Experiments show that Lapstyle produces more appealing stylized images with less artifacts, without compromising their "stylishness".

Word embedding maps words into a lowdimensional continuous embedding space by exploiting the local word collocation patterns in a small context window.On the other hand, topic modeling maps documents onto a low-dimensional topic space, by utilizing the global word collocation patterns in the same document.These two types of patterns are complementary.In this paper, we propose a generative topic embedding model to combine the two types of patterns.In our model, topics are represented by embedding vectors, and are shared across documents.The probability of each word is influenced by both its local context and its topic.A variational inference method yields the topic embeddings as well as the topic mixing proportions for each document.Jointly they represent the document in a low-dimensional continuous space.In two document classification tasks, our method performs better than eight existing methods, with fewer features.In addition, we illustrate with an example that our method can generate coherent topics even based on only one document.

Most existing word embedding methods can be categorized into Neural Embedding Models and Matrix Factorization (MF)based methods.However some models are opaque to probabilistic interpretation, and MF-based methods, typically solved using Singular Value Decomposition (SVD), may incur loss of corpus information.In addition, it is desirable to incorporate global latent factors, such as topics, sentiments or writing styles, into the word embedding model.Since generative models provide a principled way to incorporate latent factors, we propose a generative word embedding model, which is easy to interpret, and can serve as a basis of more sophisticated latent factor models.The model inference reduces to a low rank weighted positive semidefinite approximation problem.Its optimization is approached by eigendecomposition on a submatrix, followed by online blockwise regression, which is scalable and avoids the information loss in SVD.In experiments on 7 common benchmark datasets, our vectors are competitive to word2vec, and better than other MF-based methods.