Models, code, and papers for "Kai Chen":

Automatic photo cropping is an important tool for improving visual quality of digital photos without resorting to tedious manual selection. Traditionally, photo cropping is accomplished by determining the best proposal window through visual quality assessment or saliency detection. In essence, the performance of an image cropper highly depends on the ability to correctly rank a number of visually similar proposal windows. Despite the ranking nature of automatic photo cropping, little attention has been paid to learning-to-rank algorithms in tackling such a problem. In this work, we conduct an extensive study on traditional approaches as well as ranking-based croppers trained on various image features. In addition, a new dataset consisting of high quality cropping and pairwise ranking annotations is presented to evaluate the performance of various baselines. The experimental results on the new dataset provide useful insights into the design of better photo cropping algorithms.

In this paper, we point out a fundamental property of the objective in reinforcement learning, with which we can reformulate the policy gradient objective into a perceptron-like loss function, removing the need to distinguish between on and off policy training. Namely, we posit that it is sufficient to only update a policy $\pi$ for cases that satisfy the condition $A(\frac{\pi}{\mu}-1)\leq0$, where $A$ is the advantage, and $\mu$ is another policy. Furthermore, we show via theoretic derivation that a perceptron-like loss function matches the clipped surrogate objective for PPO. With our new formulation, the policies $\pi$ and $\mu$ can be arbitrarily apart in theory, effectively enabling off-policy training. To examine our derivations, we can combine the on-policy PPO clipped surrogate (which we show to be equivalent with one instance of the new reformation) with the off-policy IMPALA method. We first verify the combined method on the OpenAI Gym pendulum toy problem. Next, we use our method to train a quadrotor position controller in a simulator. Our trained policy is efficient and lightweight enough to perform in a low cost micro-controller at a minimum update rate of 500 Hz. For the quadrotor, we show two experiments to verify our method and demonstrate performance: 1) hovering at a fixed position, and 2) tracking along a specific trajectory. In preliminary trials, we are also able to apply the method to a real-world quadrotor.

There are several challenges for search and rescue robots: mobility, perception, autonomy, and communication. Inspired by the DARPA Subterranean (SubT) Challenge, we propose an autonomous blimp robot, which has the advantages of low power consumption and collision-tolerance compared to other aerial vehicles like drones. This is important for search and rescue tasks that usually last for one or more hours. However, the underground constrained passages limit the size of blimp envelope and its payload, making the proposed system resource-constrained. Therefore, a careful design consideration is needed to build a blimp system with on-board artifact search and SLAM. In order to reach long-term operation, a failure-aware algorithm with minimal communication to human supervisor to have situational awareness and send control signals to the blimp when needed.

This paper presents a Convolutional Neural Network (CNN) based page segmentation method for handwritten historical document images. We consider page segmentation as a pixel labeling problem, i.e., each pixel is classified as one of the predefined classes. Traditional methods in this area rely on carefully hand-crafted features or large amounts of prior knowledge. In contrast, we propose to learn features from raw image pixels using a CNN. While many researchers focus on developing deep CNN architectures to solve different problems, we train a simple CNN with only one convolution layer. We show that the simple architecture achieves competitive results against other deep architectures on different public datasets. Experiments also demonstrate the effectiveness and superiority of the proposed method compared to previous methods.

Recently, discriminatively learned correlation filters (DCF) has drawn much attention in visual object tracking community. The success of DCF is potentially attributed to the fact that a large amount of samples are utilized to train the ridge regression model and predict the location of object. To solve the regression problem in an efficient way, these samples are all generated by circularly shifting from a search patch. However, these synthetic samples also induce some negative effects which weaken the robustness of DCF based trackers. In this paper, we propose a Convolutional Regression framework for visual tracking (CRT). Instead of learning the linear regression model in a closed form, we try to solve the regression problem by optimizing a one-channel-output convolution layer with Gradient Descent (GD). In particular, the receptive field size of the convolution layer is set to the size of object. Contrary to DCF, it is possible to incorporate all "real" samples clipped from the whole image. A critical issue of the GD approach is that most of the convolutional samples are negative and the contribution of positive samples will be suppressed. To address this problem, we propose a novel Automatic Hard Negative Mining method to eliminate easy negatives and enhance positives. Extensive experiments are conducted on a widely-used benchmark with 100 sequences. The results show that the proposed algorithm achieves outstanding performance and outperforms almost all the existing DCF based algorithms.

One of the main challenges of visual object tracking comes from the arbitrary appearance of objects. Most existing algorithms try to resolve this problem as an object-specific task, i.e., the model is trained to regenerate or classify a specific object. As a result, the model need to be initialized and retrained for different objects. In this paper, we propose a more generic approach utilizing a novel two-flow convolutional neural network (named YCNN). The YCNN takes two inputs (one is object image patch, the other is search image patch), then outputs a response map which predicts how likely the object appears in a specific location. Unlike those object-specific approach, the YCNN is trained to measure the similarity between two image patches. Thus it will not be confined to any specific object. Furthermore the network can be end-to-end trained to extract both shallow and deep convolutional features which are dedicated for visual tracking. And once properly trained, the YCNN can be applied to track all kinds of objects without further training and updating. Benefiting from the once-for-all model, our algorithm is able to run at a very high speed of 45 frames-per-second. The experiments on 51 sequences also show that our algorithm achieves an outstanding performance.

The performance on deep learning is significantly affected by volume of training data. Models pre-trained from massive dataset such as ImageNet become a powerful weapon for speeding up training convergence and improving accuracy. Similarly, models based on large dataset are important for the development of deep learning in 3D medical images. However, it is extremely challenging to build a sufficiently large dataset due to difficulty of data acquisition and annotation in 3D medical imaging. We aggregate the dataset from several medical challenges to build 3DSeg-8 dataset with diverse modalities, target organs, and pathologies. To extract general medical three-dimension (3D) features, we design a heterogeneous 3D network called Med3D to co-train multi-domain 3DSeg-8 so as to make a series of pre-trained models. We transfer Med3D pre-trained models to lung segmentation in LIDC dataset, pulmonary nodule classification in LIDC dataset and liver segmentation on LiTS challenge. Experiments show that the Med3D can accelerate the training convergence speed of target 3D medical tasks 2 times compared with model pre-trained on Kinetics dataset, and 10 times compared with training from scratch as well as improve accuracy ranging from 3% to 20%. Transferring our Med3D model on state-the-of-art DenseASPP segmentation network, in case of single model, we achieve 94.6\% Dice coefficient which approaches the result of top-ranged algorithms on the LiTS challenge.

With superiorities on low cost, portability, and free of radiation, echocardiogram is a widely used imaging modality for left ventricle (LV) function quantification. However, automatic LV segmentation and motion tracking is still a challenging task. In addition to fuzzy border definition, low contrast, and abounding artifacts on typical ultrasound images, the shape and size of the LV change significantly in a cardiac cycle. In this work, we propose a temporal affine network (TAN) to perform image analysis in a warped image space, where the shape and size variations due to the cardiac motion as well as other artifacts are largely compensated. Furthermore, we perform three frequent echocardiogram interpretation tasks simultaneously: standard cardiac plane recognition, LV landmark detection, and LV segmentation. Instead of using three networks with one dedicating to each task, we use a multi-task network to perform three tasks simultaneously. Since three tasks share the same encoder, the compact network improves the segmentation accuracy with more supervision. The network is further finetuned with optical flow adjusted annotations to enhance motion coherence in the segmentation result. Experiments on 1,714 2D echocardiographic sequences demonstrate that the proposed method achieves state-of-the-art segmentation accuracy with real-time efficiency.

Local misalignment caused by global homography is a common issue in image stitching task. Content-Preserving Warping (CPW) is a typical method to deal with this issue, in which geometric and photometric constraints are imposed to guide the warping process. One of its essential condition however, is colour consistency, and an elusive goal in real world applications. In this paper, we propose a Generalized Content-Preserving Warping (GCPW) method to alleviate this problem. GCPW extends the original CPW by applying a colour model that expresses the colour transformation between images locally, thus meeting the photometric constraint requirements for effective image stitching. We combine the photometric and geometric constraints and jointly estimate the colour transformation and the warped mesh vertexes, simultaneously. We align images locally with an optimal grid mesh generated by our GCPW method. Experiments on both synthetic and real images demonstrate that our new method is robust to colour variations, outperforming other state-of-the-art CPW-based image stitching methods.

Speech recognition is a sequence prediction problem. Besides employing various deep learning approaches for framelevel classification, sequence-level discriminative training has been proved to be indispensable to achieve the state-of-the-art performance in large vocabulary continuous speech recognition (LVCSR). However, keyword spotting (KWS), as one of the most common speech recognition tasks, almost only benefits from frame-level deep learning due to the difficulty of getting competing sequence hypotheses. The few studies on sequence discriminative training for KWS are limited for fixed vocabulary or LVCSR based methods and have not been compared to the state-of-the-art deep learning based KWS approaches. In this paper, a sequence discriminative training framework is proposed for both fixed vocabulary and unrestricted acoustic KWS. Sequence discriminative training for both sequence-level generative and discriminative models are systematically investigated. By introducing word-independent phone lattices or non-keyword blank symbols to construct competing hypotheses, feasible and efficient sequence discriminative training approaches are proposed for acoustic KWS. Experiments showed that the proposed approaches obtained consistent and significant improvement in both fixed vocabulary and unrestricted KWS tasks, compared to previous frame-level deep learning based acoustic KWS methods.

In the domain of pattern recognition, using the SPD (Symmetric Positive Definite) matrices to represent data and taking the metrics of resulting Riemannian manifold into account have been widely used for the task of image set classification. In this paper, we propose a new data representation framework for image sets named CSPD (Component Symmetric Positive Definite). Firstly, we obtain sub-image sets by dividing the image set into square blocks with the same size, and use traditional SPD model to describe them. Then, we use the results of the Riemannian kernel on SPD matrices as similarities of corresponding sub-image sets. Finally, the CSPD matrix appears in the form of the kernel matrix for all the sub-image sets, and CSPDi,j denotes the similarity between i-th sub-image set and j-th sub-image set. Here, the Riemannian kernel is shown to satisfy the Mercer's theorem, so our proposed CSPD matrix is symmetric and positive definite and also lies on a Riemannian manifold. On three benchmark datasets, experimental results show that CSPD is a lower-dimensional and more discriminative data descriptor for the task of image set classification.

There are many situations for which an unmanned ground vehicle has to work with only partial observability of the environment. Therefore, a feasible nonholonomic obstacle avoidance and target tracking action must be generated immediately based on the real-time perceptual information. This paper presents a robust approach to integrating VPH+ (enhanced vector polar histogram) and MPC (model predictive control). VPH+ is applied to calculate the desired direction for its environment perception ability and computational efficiency, while MPC is explored to perform a constrained model-predictive trajectory generation. This approach can be implemented in a reactive controller. Simulation experiments are performed in VREP to validate the proposed approach.

We investigate the scenario that a robot needs to reach a designated goal after taking a sequence of appropriate actions in a non-static environment that is partially structured. One application example is to control a marine vehicle to move in the ocean. The ocean environment is dynamic and oftentimes the ocean waves result in strong disturbances that can disturb the vehicle's motion. Modeling such dynamic environment is non-trivial, and integrating such model in the robotic motion control is particularly difficult. Fortunately, the ocean currents usually form some local patterns (e.g. vortex) and thus the environment is partially structured. The historically observed data can be used to train the robot to learn to interact with the ocean tidal disturbances. In this paper we propose a method that applies the deep reinforcement learning framework to learn such partially structured complex disturbances. Our results show that, by training the robot under artificial and real ocean disturbances, the robot is able to successfully act in complex and spatiotemporal environments.

Path planning is an important problem in robotics. One way to plan a path between two points $x,y$ within a (not necessarily simply-connected) planar domain $\Omega$, is to define a non-negative distance function $d(x,y)$ on $\Omega\times\Omega$ such that following the (descending) gradient of this distance function traces such a path. This presents two equally important challenges: A mathematical challenge -- to define $d$ such that $d(x,y)$ has a single minimum for any fixed $y$ (and this is when $x=y$), since a local minimum is in effect a "dead end", A computational challenge -- to define $d$ such that it may be computed efficiently. In this paper, given a description of $\Omega$, we show how to assign coordinates to each point of $\Omega$ and define a family of distance functions between points using these coordinates, such that both the mathematical and the computational challenges are met. This is done using the concepts of \emph{harmonic measure} and \emph{$f$-divergences}. In practice, path planning is done on a discrete network defined on a finite set of \emph{sites} sampled from $\Omega$, so any method that works well on the continuous domain must be adapted so that it still works well on the discrete domain. Given a set of sites sampled from $\Omega$, we show how to define a network connecting these sites such that a \emph{greedy routing} algorithm (which is the discrete equivalent of continuous gradient descent) based on the distance function mentioned above is guaranteed to generate a path in the network between any two such sites. In many cases, this network is close to a (desirable) planar graph, especially if the set of sites is dense.

Distance functions between points in a domain are sometimes used to automatically plan a gradient-descent path towards a given target point in the domain, avoiding obstacles that may be present. A key requirement from such distance functions is the absence of spurious local minima, which may foil such an approach, and this has led to the common use of harmonic potential functions. Based on the planar Laplace operator, the potential function guarantees the absence of spurious minima, but is well known to be slow to numerically compute and prone to numerical precision issues. To alleviate the first of these problems, we propose a family of novel divergence distances. These are based on f-divergence of the Poisson kernel of the domain. We define the divergence distances and compare them to the harmonic potential function and other related distance functions. Our first result is theoretical: We show that the family of divergence distances are equivalent to the harmonic potential function on simply-connected domains, namely generate paths which are identical to those generated by the potential function. The proof is based on the concept of conformal invariance. Our other results are more practical and relate to two special cases of divergence distances, one based on the Kullback-Leibler divergence and one based on the total variation divergence. We show that using divergence distances instead of the potential function and other distances has a significant computational advantage, as, following a pre-processing stage, they may be computed up to an order of magnitude faster than the others when taking advantage of certain sparsity properties of the Poisson kernel. Furthermore, the computation is "embarrassingly parallel", so may be implemented on a GPU with up to three orders of magnitude speedup.

Candidate text region extraction plays a critical role in convolutional neural network (CNN) based text detection from natural images. In this paper, we propose a CNN based scene text detection algorithm with a new text region extractor. The so called candidate text region extractor I-MSER is based on Maximally Stable Extremal Region (MSER), which can improve the independency and completeness of the extracted candidate text regions. Design of I-MSER is motivated by the observation that text MSERs have high similarity and are close to each other. The independency of candidate text regions obtained by I-MSER is guaranteed by selecting the most representative regions from a MSER tree which is generated according to the spatial overlapping relationship among the MSERs. A multi-layer CNN model is trained to score the confidence value of the extracted regions extracted by the I-MSER for text detection. The new text detection algorithm based on I-MSER is evaluated with wide-used ICDAR 2011 and 2013 datasets and shows improved detection performance compared to the existing algorithms.

Scene text recognition plays an important role in many computer vision applications. The small size of available public available scene text datasets is the main challenge when training a text recognition CNN model. In this paper, we propose a CNN based Chinese text recognition algorithm. To enlarge the dataset for training the CNN model, we design a synthetic data engine for Chinese scene character generation, which generates representative character images according to the fonts use frequency of Chinese texts. As the Chinese text is more complex, the English text recognition CNN architecture is modified for Chinese text. To ensure the small size nature character dataset and the large size artificial character dataset are comparable in training, the CNN model are trained progressively. The proposed Chinese text recognition algorithm is evaluated with two Chinese text datasets. The algorithm achieves better recognize accuracy compared to the baseline methods.

Extended Kalman filter (EKF) does not guarantee consistent mean and covariance under linearization, even though it is the main framework for robotic localization. While Lie group improves the modeling of the state space in localization, the EKF on Lie group still relies on the arbitrary Gaussian assumption in face of nonlinear models. We instead use von Mises filter for orientation estimation together with the conventional Kalman filter for position estimation, and thus we are able to characterize the first two moments of the state estimates. Since the proposed algorithm holds a solid probabilistic basis, it is fundamentally relieved from the inconsistency problem. Furthermore, we extend the localization algorithm to fully circular representation even for position, which is similar to grid patterns found in mammalian brains and in recurrent neural networks. The applicability of the proposed algorithms is substantiated not only by strong mathematical foundation but also by the comparison against other common localization methods.

Spectral Mixture (SM) kernels form a powerful class of kernels for Gaussian processes, capable to discover patterns, extrapolate, and model negative co-variances. In SM kernels, spectral mixture components are linearly combined to construct a final flexible kernel. As a consequence SM kernels does not explicitly model correlations between components and dependencies related to time and phase delays between components, because only the auto-convolution of base components are used. To address these drawbacks we introduce Generalized Convolution Spectral Mixture (GCSM) kernels. We incorporate time and phase delay into the base spectral mixture and use cross-convolution between a base component and the complex conjugate of another base component to construct a complex-valued and positive definite kernel representing correlations between base components. In this way the total number of components in GCSM becomes quadratic. We perform a thorough comparative experimental analysis of GCSM on synthetic and real-life datasets. Results indicate the beneficial effect of the extra features of GCSM. This is illustrated in the problem of forecasting the long range trend of a river flow to monitor environment evolution, where GCSM is capable of discovering correlated patterns that SM cannot and improving patterns recognition ability of SM.

Multi-output Gaussian processes (MOGPs) are recently extended by using spectral mixture kernel, which enables expressively pattern extrapolation with a strong interpretation. In particular, Multi-Output Spectral Mixture kernel (MOSM) is a recent, powerful state of the art method. However, MOSM cannot reduce to the ordinary spectral mixture kernel (SM) when using a single channel. Moreover, when the spectral density of different channels is either very close or very far from each other in the frequency domain, MOSM generates unreasonable scale effects on cross weights which produces an incorrect description of the channel correlation structure. In this paper, we tackle these drawbacks and introduce a principled multi-output convolution spectral mixture kernel (MOCSM) framework. In our framework, we model channel dependencies through convolution of time and phase delayed spectral mixtures between different channels. Results of extensive experiments on synthetic and real datasets demontrate the advantages of MOCSM and its state of the art performance.