Parameter Learning and Convergent Inference for Dense Random Fields

By Philipp Krahenbuhl, Vladlen Koltun. 2013

238 citations recorded by the source corpus.

Subfields: Probabilistic & Bayesian Methods

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Probabilistic & Bayesian Methods

Citations in the corpus

Pixelwise Instance Segmentation with a Dynamically Instantiated Network (2017, neutral): “We found that this performed better than the approximate IoU loss proposed in [24, 38].”
Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs (2014, methodology): “There are multiple aspects in our model that we intend to reﬁne, such as fully integrating its two main components (CNN and CRF) and train the whole system in an end-to-end fashion, similar to Kr¨ahenb¨uhl & Koltun (2013); Chen et al. (2014); Zheng et al. (2015).”
Conditional Random Fields as Recurrent Neural Networks (2015, methodology): “Loss function During the training of the models that achieved the best results reported in this paper, we used the standard softmax loss function, that is, the log-likelihood error function described in [30].”
Efficient Piecewise Training of Deep Structured Models for Semantic Segmentation (2015, methodology): “The method in [31] extends DeepLab and [22] by jointly learning the dense CRFs and CNNs.”
Fully Connected Deep Structured Networks (2015, methodology): “We refer the interested reader to [18] for additional details regarding the computation of the gradient ∂q(x,y),i(ŷi) ∂w .”
Semantic Image Segmentation via Deep Parsing Network (2015, background): “wise functions, including long-range dependencies [16, 17], high-order potentials [37, 36], and semantic label contexts [21, 26, 38].”
Material recognition in the wild with the Materials in Context Database (2014, methodology): “Since there are only a few parameters to solve ( θ p , θ ab , θ L , w p ), we select different parameter settings of the dense CRF and different ensemble sizes of CNNs to average, resulting in 465 different methods to consider.”
Higher Order Conditional Random Fields in Deep Neural Networks (2015, background): “Some early works that advocated gradient backpropagation through graphical model inference for parameter optimisation include [32, 33] and [34].”
Semantic Segmentation using Adversarial Networks (2016, methodology): “For certain classes of pairwise potentials, mean-field inference in fully-connected CRFs with millions of variables is tractable using recent filter-based techniques[14].”
Pixel-Adaptive Convolutional Neural Networks (2019, methodology): “Previous work [24, 25] relies on permutohedral lattice convolution [1] to achieve efﬁcient implementation.”
DeepIGeoS: A Deep Interactive Geodesic Framework for Medical Image Segmentation (2017, methodology): “Other methods including structured output Support Vector Machines [47], approximate marginal inference [48] and gradient-based optimization [49] were also proposed to learn parameters in CRFs.”
Higher Order Potentials in End-to-End Trainable Conditional Random Fields (2015, background): “Some early works that advocated gradient backpropagation through graphical model inference for parameter optimization include [35, 8, 22] and [17].”
Large-Scale 3D Shape Reconstruction and Segmentation from ShapeNet Core55 (2017, methodology): “We redesign a Pointnet [13] structure as a dense connection block like DenseNet [8] and After training, we use the dense conditional random ﬁeld (CRF) [11]to reﬁne the results.”
On Regularized Losses for Weakly-supervised CNN Segmentation (2018, background): “As hyper-parameter, the Gaussian bandwidth is optimized via validation for DenseCRF, normalized cut and kernel cut.”
A Review on Deep Learning Techniques Applied to Semantic Segmentation (2017, methodology): “The DeepLab models [68] [69] make use of the fully connected pairwise CRF by Krähenbühl and Koltun [94] [95] as a separated post-processing step in their pipeline to refine the segmentation result.”
Training Deep Networks with Structured Layers by Matrix Backpropagation (2015, methodology): “[37] obtained state-of-the-art semantic segmentation results using an architecture similar to [36] but enforcing structure using globally connected CRFs[38] where only the unary potentials are learnt.”