Regularized Loss With Hyperparameter Estimation for Weakly Supervised Single Class Segmentation.

We propose a new image level weakly supervised segmentation approach for datasets with a single object class of interest. Our approach is based on a regularized loss function inspired by the classical Conditional Random Field (CRF) modeling. Our loss models properties of generic objects, and we use it to guide CNN towards segments that are more likely to correspond to the object, thus avoiding the need for pixel precise annotations. Training CNN with regularized loss is a difficult task for gradient descent. We develop an annealing algorithm which is crucial for a successful training. Furthermore, we develop an approach for hyperparameter setting for the most important components of our regularized loss. This is far from trivial, since there is no pixel precise ground truth for guidance. The advantage of our method is that we use a standard CNN architecture and an easy to interpret loss function, derived from classical CRF models. Furthermore, we apply the same loss function for any task/dataset. We first evaluate our approach for salient object segmentation and co-segmentation. These tasks naturally involve one object class of interest. Then we adapt our approach to image level weakly supervised multi-class semantic segmentation. We obtain state-of-the-art results.

Efficient Graph Cut Optimization for Full CRFs with Quantized Edges.

Fully connected pairwise Conditional Random Fields (Full-CRF) with Gaussian edge weights can achieve superior results compared to sparsely connected CRFs. However, traditional methods for Full-CRFs are too expensive. Previous work develops efficient approximate optimization based on mean field inference, which is a local optimization method and can be far from the optimum. We propose efficient and effective optimization based on graph cuts for Full-CRFs with quantized edge weights. To quantize edge weights, we partition the image into superpixels and assume that the weight of an edge between any two pixels depends only on the superpixels these pixels belong to. Our quantized edge CRF is an approximation to the Gaussian edge CRF, and gets closer to it as superpixel size decreases. Being an approximation, our model offers an intuition about the regularization properties of the Guassian edge Full-CRF. For efficient inference, we first consider the two-label case and develop an approximate method based on transforming the original problem into a smaller domain. Then we handle multi-label CRF by showing how to implement expansion moves. In both binary and multi-label cases, our solutions have significantly lower energy compared to that of mean field inference. We also show the effectiveness of our approach on semantic segmentation task.

Convexity Shape Prior for Binary Segmentation.

Convexity is a known important cue in human vision. We propose shape convexity as a new high-order regularization constraint for binary image segmentation. In the context of discrete optimization, object convexity is represented as a sum of three-clique potentials penalizing any 1- 0- 1 configuration on all straight lines. We show that these non-submodular potentials can be efficiently optimized using an iterative trust region approach. At each iteration the energy is linearly approximated and globally optimized within a small trust region around the current solution. While the quadratic number of all three-cliques is prohibitively high, we design a dynamic programming technique for evaluating and approximating these cliques in linear time. We also derive a second order approximation model that is more accurate but computationally intensive. We discuss limitations of our local optimization and propose gradual non-submodularization scheme that alleviates some limitations. Our experiments demonstrate general usefulness of the proposed convexity shape prior on synthetic and real image segmentation examples. Unlike standard second-order length regularization, our convexity prior does not have shrinking bias, and is robust to changes in scale and parameter selection.

Local Submodularization for Binary Pairwise Energies.

Many computer vision problems require optimization of binary non-submodular energies. We propose a general optimization framework based on local submodular approximations (LSA). Unlike standard LP relaxation methods that linearize the whole energy globally, our approach iteratively approximates the energy locally. On the other hand, unlike standard local optimization methods (e.g., gradient descent or projection techniques) we use non-linear submodular approximations and optimize them without leaving the domain of integer solutions. We discuss two specific LSA algorithms based on trust region and auxiliary function principles, LSA-TR and LSA-AUX. The proposed methods obtain state-of-the-art results on a wide range of applications such as binary deconvolution, curvature regularization, inpainting, segmentation with repulsion and two types of shape priors. Finally, we discuss a move-making extension to the LSA-TR approach. While our paper is focused on pairwise energies, our ideas extend to higher-order problems. The code is available online.

Prostate histopathology: learning tissue component histograms for cancer detection and classification.

Radical prostatectomy is performed on approximately 40% of men with organ-confined prostate cancer. Pathologic information obtained from the prostatectomy specimen provides important prognostic information and guides recommendations for adjuvant treatment. The current pathology protocol in most centers involves primarily qualitative assessment. In this paper, we describe and evaluate our system for automatic prostate cancer detection and grading on hematoxylin & eosin-stained tissue images. Our approach is intended to address the dual challenges of large data size and the need for high-level tissue information about the locations and grades of tumors. Our system uses two stages of AdaBoost-based classification. The first provides high-level tissue component labeling of a superpixel image partitioning. The second uses the tissue component labeling to provide a classification of cancer versus noncancer, and low-grade versus high-grade cancer. We evaluated our system using 991 sub-images extracted from digital pathology images of 50 whole-mount tissue sections from 15 prostatectomy patients. We measured accuracies of 90% and 85% for the cancer versus noncancer and high-grade versus low-grade classification tasks, respectively. This system represents a first step toward automated cancer quantification on prostate digital histopathology imaging, which could pave the way for more accurately informed postprostatectomy patient care.

Interactive Segmentation with Super-Labels.

In interactive segmentation, the most common way to model object appearance is by GMM or histogram, while MRFs are used to encourage spatial coherence among the object labels. This makes the strong assumption that pixels within each object are i.i.d. when in fact most objects have multiple distinct appearances and exhibit strong spatial correlation among their pixels. At the very least, this calls for an MRF-based appearance model within each object itself and yet, to the best of our knowledge, such a "two-level MRF" has never been proposed. We propose a novel segmentation energy that can model complex appearance. We represent the appearance of each object by a set of distinct spatially coherent models. This results in a two-level MRF with "super-labels" at the top level that are partitioned into "sub-labels" at the bottom. We introduce the hierarchical Potts (hPotts) prior to govern spatial coherence within each level. Finally, we introduce a novel algorithm with EM-style alternation of proposal, α-expansion and re-estimation steps. Our experiments demonstrate the conceptual and qualitative improvement that a two-level MRF can provide. We show applications in binary segmentation, multi-class segmentation, and interactive co-segmentation. Finally, our energy and algorithm have interesting interpretations in terms of semi-supervised learning.

Order-preserving moves for graph-cut-based optimization.

In the last decade, graph-cut optimization has been popular for a variety of labeling problems. Typically, graph-cut methods are used to incorporate smoothness constraints on a labeling, encouraging most nearby pixels to have equal or similar labels. In addition to smoothness, ordering constraints on labels are also useful. For example, in object segmentation, a pixel with a "car wheel" label may be prohibited above a pixel with a "car roof" label. We observe that the commonly used graph-cut \alpha-expansion move algorithm is more likely to get stuck in a local minimum when ordering constraints are used. For a certain model with ordering constraints, we develop new graph-cut moves which we call order-preserving. The advantage of order-preserving moves is that they act on all labels simultaneously, unlike \alpha-expansion. More importantly, for most labels \alpha, the set of \alpha-expansion moves is strictly smaller than the set of order-preserving moves. This helps to explain why in practice optimization with order-preserving moves performs significantly better than \alpha-expansion in the presence of ordering constraints. We evaluate order-preserving moves for the geometric class scene labeling (introduced by Hoiem et al.) where the goal is to assign each pixel a label such as "sky," "ground," etc., so ordering constraints arise naturally. In addition, we use order-preserving moves for certain simple shape priors in graph-cut segmentation, which is a novel contribution in itself.

A comparative study of energy minimization methods for Markov random fields with smoothness-based priors.

Among the most exciting advances in early vision has been the development of efficient energy minimization algorithms for pixel-labeling tasks such as depth or texture computation. It has been known for decades that such problems can be elegantly expressed as Markov random fields, yet the resulting energy minimization problems have been widely viewed as intractable. Recently, algorithms such as graph cuts and loopy belief propagation (LBP) have proven to be very powerful: for example, such methods form the basis for almost all the top-performing stereo methods. However, the tradeoffs among different energy minimization algorithms are still not well understood. In this paper we describe a set of energy minimization benchmarks and use them to compare the solution quality and running time of several common energy minimization algorithms. We investigate three promising recent methods graph cuts, LBP, and tree-reweighted message passing in addition to the well-known older iterated conditional modes (ICM) algorithm. Our benchmark problems are drawn from published energy functions used for stereo, image stitching, interactive segmentation, and denoising. We also provide a general-purpose software interface that allows vision researchers to easily switch between optimization methods. Benchmarks, code, images, and results are available at http://vision.middlebury.edu/MRF/.