Robust Optimization for Deep Regression

By Vasileios Belagiannis, Christian Rupprecht, Gustavo Carneiro, Nassir Navab. 2015

192 citations recorded by the source corpus.

Convolutional Neural Networks (ConvNets) have successfully contributed to improve the accuracy of regression-based methods for computer vision tasks such as human pose estimation, landmark localization, and object detection. The network optimization has been usually performed with L2 loss and without considering the impact of outliers on the training process, where an outlier in this context is defined by a sample estimation that lies at an abnormal distance from the other training sample estimations in the objective space. In this work, we propose a regression model with ConvNets that achieves robustness to such outliers by minimizing Tukey's biweight function, an M-estimator robust to outliers, as the loss function for the ConvNet. In addition to the robust loss, we introduce a coarse-to-fine model, which processes input images of progressively higher resolutions for improving the accuracy of the regressed values. In our experiments, we demonstrate faster convergence and better generalization of our robust loss function for the tasks of human pose estimation and age estimation from face images. We also show that the combination of the robust loss function with the coarse-to-fine model produces comparable or better results than current state-of-the-art approaches in four publicly available human pose estimation datasets.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

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

In Computer Vision and Pattern Recognition

Citations in the corpus

2D/3D Pose Estimation and Action Recognition Using Multitask Deep Learning (2018, background): “The problem of human pose estimation has been intensively studied in the last years, from Pictorial Structures [2, 17, 37] to more recent CNN approaches [34, 25, 38, 20, 41, 54, 5, 51, 52, 36].”
Recurrent Human Pose Estimation (2016, methodology): “the hand and limb detectors, and the pairwise part detectors have become stronger, so the importance of the configuration constraints has begun to wane, with quite recent methods not even including an explicit graphical model [4],”
Deeper Depth Prediction with Fully Convolutional Residual Networks (2016, background): “Empirically, BerHu shows a good balance between the two norms in the given problem; it puts high weight towards samples/pixels with a high residual because of the L2 term, contrary for example to a robust loss, such as Tukey’s biweight function that ignores samples with high residuals [2].”
Geometric Loss Functions for Camera Pose Regression with Deep Learning (2017, methodology): “We can also use robust norms such as Huber’s loss [16] and Tukey’s loss [15], which have been successfully applied to deep learning [5].”
A Comprehensive Analysis of Deep Regression (2018, unknown): “Such lines are drawn by computing the difference in performance between the best performing method in Tables 13, 11 and 12 for Biwi, FLD and Parse, respectively, and the second best performing method in the corresponding original papers [23], [18], and [19].”
Learning to Count with CNN Boosting (2016, methodology): “In addition, we evaluated the sample selection scheme proposed in [32], which employs a robust loss function based on Tukeys biweight function that weighs the training samples based on the residual magnitude.”
Human Pose Regression by Combining Indirect Part Detection and Contextual Information (2017, unknown): “More recent methods have shown signiﬁcant improvements on accuracy by using fully convolutional models to generate belief maps (or heat maps) for joint probabilities [34, 3, 5, 27, 13].”
Neural Body Fitting: Unifying Deep Learning and Model Based Human Pose and Shape Estimation (2018, methodology): “advent of robust 2D pose methods – by matching 2D detections to a 3D pose database [8,66], by regressing pose from 2D joint distance matrices [35], by exploiting pose and geometric priors for lifting [69,1,51,19,32,70,47]; or simply by training a feed forward network to directly predict 3D pose from”
Deep Label Distribution Learning With Label Ambiguity (2016, background): “Similar difficulties arise in head pose estimation, where head poses are usually collected at a small set of angles with a 10◦ or 15◦ increment.”
Fashion Landmark Detection in the Wild (2016, background): “Recently, deep learning [10,11,12,29] has shown great advantages in locating human joints and there are generally two directions here.”
Hough-CNN: Deep learning for segmentation of deep brain regions in MRI and ultrasound (2016, background): “Classification problems such as image categorisation [1], [2], object detection [23] and face recognition [24] as well as regression problems such as human pose estimation [25], and depth prediction from RGB data [26] have been addressed using CNNs and unprecedented results have been presented to th”
Revisiting Projection-Free Optimization for Strongly Convex Constraint Sets (2018, methodology): “For robust regression, we used the bi-weight loss (Belagiannis et al. 2015), as a non-convex (but smooth) loss ψ ( f For regression, we used the YearPredictionMSD dataset (500K observations, 90 features) (Lichman 2013).”
Human Pose Estimation via Convolutional Part Heatmap Regression (2016, methodology): “Recently proposed methods for articulated human pose estimation using CNNs can be classiﬁed as detection-based [22–26] or regression-based [14–17, 27, 28].”
Knowledge-Guided Deep Fractal Neural Networks for Human Pose Estimation (2017, methodology): “CNN features are regressed in order to provide joint predictions of the body parts [10], [16], [42].”
AggNet: Deep Learning From Crowds for Mitosis Detection in Breast Cancer Histology Images (2016, methodology): “Indeed, deep learning has advanced the field of computer vision the last few years [19] leading to powerful methods for various applications such as object classification [5], detection [20], segmentation [21], robust regression [22] and depth prediction [23].”
Deep-Learned Collision Avoidance Policy for Distributed Multiagent Navigation (2016, background): “function for regression tasks usually is more fragile to outliers [23].”