Higher Order Conditional Random Fields Research Articles

Higher-Order Conditional Random Field established with CNNs for Video Object Segmentation

Higher-Order Conditional Random Field established with CNNs for Video Object Segmentation

Incorporating DeepLabv3+ and object-based image analysis for semantic segmentation of very high resolution remote sensing images

ABSTRACT Semantic segmentation of remote sensing images is an important but unsolved problem in the remote sensing society. Advanced image semantic segmentation models, such as DeepLabv3+, have achieved astonishing performance for semantically labeling very high resolution (VHR) remote sensing images. However, it is difficult for these models to capture the precise outlines of ground objects and explore the context information that revealing relationships among image objects for optimizing segmentation results. Consequently, this study proposes a semantic segmentation method for VHR images by incorporating deep learning semantic segmentation model (DeepLabv3+) and object-based image analysis (OBIA), wherein DSM is employed to provide geometric information to enhance the interpretation of VHR images. The proposed method first obtains two initial probabilistic labeling predictions using a DeepLabv3+ network on spectral image and a random forest (RF) classifier on hand-crafted features, respectively. These two predictions are then integrated by Dempster-Shafer (D-S) evidence theory to be fed into an object-constrained higher-order conditional random field (CRF) framework to estimate the final semantic labeling results with the consideration of the spatial contextual information. The proposed method is applied to the ISPRS 2D semantic labeling benchmark, and competitive overall accuracies of 90.6% and 85.0% are achieved for Vaihingen and Potsdam datasets, respectively.

Higher-Order Conditional Random Fields-Based 3D Semantic Labeling of Airborne Laser-Scanning Point Clouds

This paper presents a novel framework to achieve 3D semantic labeling of objects (e.g., trees, buildings, and vehicles) from airborne laser-scanning point clouds. To this end, we propose a framework which consists of hierarchical clustering and higher-order conditional random fields (CRF) labeling. In the hierarchical clustering, the raw point clouds are over-segmented into a set of fine-grained clusters by integrating the point density clustering and the classic K-means clustering algorithm, followed by the proposed probability density clustering algorithm. Through this process, we not only obtain a more uniform size and more homogeneous clusters with semantic consistency, but the topological relationships of the cluster’s neighborhood are implicitly maintained by turning the problem of topology maintenance into a clustering problem based on the proposed probability density clustering algorithm. Subsequently, the fine-grained clusters and their topological context are fed into the CRF labeling step, from which the fine-grained cluster’s semantic labels are learned and determined by solving a multi-label energy minimization formulation, which simultaneously considers the unary, pairwise, and higher-order potentials. Our experiments of classifying urban and residential scenes demonstrate that the proposed approach reaches 88.5% and 86.1% of “m F 1 ” estimated by averaging all classes of the F 1 -scores. We prove that the proposed method outperforms five other state-of-the-art methods. In addition, we demonstrate the effectiveness of the proposed energy terms by using an “ablation study” strategy.

Object codetection based on a higher-order conditional random field

In recent years, codetecting objects through the use of contextual information across multiple images has attracted considerable attention. We introduce an object codetection method that exploits contextual information among multiple images through a higher-order conditional random field (CRF). First, we obtain object candidates from each image of a test set by using a pretrained detector. Second, we feed the object candidates into a higher-order CRF that captures the appearance similarity using pairwise potentials and object category co-occurrence constraints using higher-order potentials. Finally, we jointly predict the category labels of all object candidates through the mean field inference in the CRF. Experimental results on the Caltech Pedestrian, PASCAL VOC 2007, PASCAL VOC 2012, and COCO datasets demonstrate the effectiveness of the proposed method compared to the baseline method.

Semantic segmentation of multisensor remote sensing imagery with deep ConvNets and higher-order conditional random fields

Aerial images acquired by multiple sensors provide comprehensive and diverse information of materials and objects within a surveyed area. The current use of pretrained deep convolutional neural networks (DCNNs) is usually constrained to three-band images (i.e., RGB) obtained from a single optical sensor. Additional spectral bands from a multiple sensor setup introduce challenges for the use of DCNN. We fuse the RGB feature information obtained from a deep learning framework with light detection and ranging (LiDAR) features to obtain semantic labeling. Specifically, we propose a decision-level multisensor fusion technique for semantic labeling of the very-high-resolution optical imagery and LiDAR data. Our approach first obtains initial probabilistic predictions from two different sources: one from a pretrained neural network fine-tuned on a three-band optical image, and another from a probabilistic classifier trained on LiDAR data. These two predictions are then combined as the unary potential using a higher-order conditional random field (CRF) framework, which resolves fusion ambiguities by exploiting the spatial–contextual information. We utilize graph cut to efficiently infer the final semantic labeling for our proposed higher-order CRF framework. Experiments performed on three benchmarking multisensor datasets demonstrate the performance advantages of our proposed method.

Learning to group and label fine-grained shape components

A majority of stock 3D models in modern shape repositories are assembled with many fine-grained components. The main cause of such data form is the component-wise modeling process widely practiced by human modelers. These modeling components thus inherently reflect some function-based shape decomposition the artist had in mind during modeling. On the other hand, modeling components represent an over-segmentation since a functional part is usually modeled as a multi-component assembly. Based on these observations, we advocate that labeled segmentation of stock 3D models should not overlook the modeling components and propose a learning solution to grouping and labeling of the fine-grained components. However, directly characterizing the shape of individual components for the purpose of labeling is unreliable, since they can be arbitrarily tiny and semantically meaningless. We propose to generate part hypotheses from the components based on a hierarchical grouping strategy, and perform labeling on those part groups instead of directly on the components. Part hypotheses are mid-level elements which are more probable to carry semantic information. A multi-scale 3D convolutional neural network is trained to extract context-aware features for the hypotheses. To accomplish a labeled segmentation of the whole shape, we formulate higher-order conditional random fields (CRFs) to infer an optimal label assignment for all components. Extensive experiments demonstrate that our method achieves significantly robust labeling results on raw 3D models from public shape repositories. Our work also contributes the first benchmark for component-wise labeling.

Protein alignment based on higher order conditional random fields for template-based modeling.

The query-template alignment of proteins is one of the most critical steps of template-based modeling methods used to predict the 3D structure of a query protein. This alignment can be interpreted as a temporal classification or structured prediction task and first order Conditional Random Fields have been proposed for protein alignment and proven to be rather successful. Some other popular structured prediction problems, such as speech or image classification, have gained from the use of higher order Conditional Random Fields due to the well known higher order correlations that exist between their labels and features. In this paper, we propose and describe the use of higher order Conditional Random Fields for query-template protein alignment. The experiments carried out on different public datasets validate our proposal, especially on distantly-related protein pairs which are the most difficult to align.

A higher order conditional random field model for simultaneous classification of land cover and land use

We propose a new approach for the simultaneous classification of land cover and land use considering spatial as well as semantic context. We apply a Conditional Random Fields (CRF) consisting of a land cover and a land use layer. In the land cover layer of the CRF, the nodes represent super-pixels; in the land use layer, the nodes correspond to objects from a geospatial database. Intra-layer edges of the CRF model spatial dependencies between neighbouring image sites. All spatially overlapping sites in both layers are connected by inter-layer edges, which leads to higher order cliques modelling the semantic relation between all land cover and land use sites in the clique. A generic formulation of the higher order potential is proposed. In order to enable efficient inference in the two-layer higher order CRF, we propose an iterative inference procedure in which the two classification tasks mutually influence each other. We integrate contextual relations between land cover and land use in the classification process by using contextual features describing the complex dependencies of all nodes in a higher order clique. These features are incorporated in a discriminative classifier, which approximates the higher order potentials during the inference procedure. The approach is designed for input data based on aerial images. Experiments are carried out on two test sites to evaluate the performance of the proposed method. The experiments show that the classification results are improved compared to the results of a non-contextual classifier. For land cover classification, the result is much more homogeneous and the delineation of land cover segments is improved. For the land use classification, an improvement is mainly achieved for land use objects showing non-typical characteristics or similarities to other land use classes. Furthermore, we have shown that the size of the super-pixels has an influence on the level of detail of the classification result, but also on the degree of smoothing induced by the segmentation method, which is especially beneficial for land cover classes covering large, homogeneous areas.

An Optimized Higher Order CRF for Automated Labeling and Segmentation of Video Objects

In this paper, we propose an optimized higher order conditional random field (CRF) labeling approach toward automated video object segmentation. Our approach introduces a computerized optimization scheme to fine tune the CRF-associated parameters, and hence make the labeling of segmented regions optimal in formulating the video objects. In comparison with the existing efforts using CRF, our optimized CRF has introduced a number of novel features, which can be highlighted as: 1) higher order CRF labeling is made adaptive to video content changes via a windowed dynamics; 2) fusion of multiple features is automatically optimized via fuzzy modeling of incoming video content and regression of parameters; 3) unary potential of higher order CRF labeling is modulated by the shortest path between neighboring regions to improve the effectiveness of higher order CRF labeling; and 4) making the algorithm affordable for a simpler graph-based video segmentation to reduce the overall computing cost, making the proposed algorithm more efficient without compromising on its performances.

Road networks as collections of minimum cost paths

We present a probabilistic representation of network structures in images. Our target application is the extraction of urban roads from aerial images. Roads appear as thin, elongated, partially curved structures forming a loopy graph, and this complex layout requires a prior that goes beyond standard smoothness and co-occurrence assumptions. In the proposed model the network is represented as a union of 1D paths connecting distant (super-)pixels. A large set of putative candidate paths is constructed in such a way that they include the true network as much as possible, by searching for minimum cost paths in the foreground (road) likelihood. Selecting the optimal subset of candidate paths is posed as MAP inference in a higher-order conditional random field. Each path forms a higher-order clique with a type of clique potential, which attracts the member nodes of cliques with high cumulative road evidence to the foreground label. That formulation induces a robust PN-Potts model, for which a global MAP solution can be found efficiently with graph cuts. Experiments with two road data sets show that the proposed model significantly improves per-pixel accuracies as well as the overall topological network quality with respect to several baselines.

Natural scene text detection with multi-layer segmentation and higher order conditional random field based analysis

Text detection in natural scene images is a hot and challenging problem in pattern recognition and computer vision. Considering the complex situations in natural scene images, we propose a robust two-steps method in this paper based on multi-layer segmentation and higher order conditional random field (CRF). Given an input image, the method separates text from its background by using multi-layer segmentation, which decomposes the input image into nine layers. Then, the connected components (CCs) in these different layers are obtained as candidate text. These candidate text CCs are verified by higher order CRF based analysis. Inspired from the multistage information integration mechanism of visual brains, features from three different levels, including separate CCs, CC pairs and CC strings, are integrated by a higher order CRF model to distinguish text from non-text. The remaining CCs are then grouped into words for easy evaluation. Experiments on the ICDAR datasets and street view dataset show that the proposed method achieves the state-of-art in natural scene text detection.

2-Slave Dual Decomposition for Generalized Higher Order CRFs

We show that the decoding problem in generalized Higher Order Conditional Random Fields (CRFs) can be decomposed into two parts: one is a tree labeling problem that can be solved in linear time using dynamic programming; the other is a supermodular quadratic pseudo-Boolean maximization problem, which can be solved in cubic time using a minimum cut algorithm. We use dual decomposition to force their agreement. Experimental results on Twitter named entity recognition and sentence dependency tagging tasks show that our method outperforms spanning tree based dual decomposition.

Robust Higher Order Potentials for Enforcing Label Consistency

This paper proposes a novel framework for labelling problems which is able to combine multiple segmentations in a principled manner. Our method is based on higher order conditional random fields and uses potentials defined on sets of pixels (image segments) generated using unsupervised segmentation algorithms. These potentials enforce label consistency in image regions and can be seen as a strict generalization of the commonly used pairwise contrast sensitive smoothness potentials. The higher order potential functions used in our framework take the form of the robust Pn model. This enables the use of powerful graph cut based move making algorithms for performing inference in the framework [14 ]. We test our method on the problem of multi-class object segmentation by augmenting the conventional CRF used for object segmentation with higher order potentials defined on image regions. Experiments on challenging data sets show that integration of higher order potentials quantitatively and qualitatively improves results leading to much better definition of object boundaries. We believe that this method can be used to yield similar improvements for many other labelling problems.