A Comprehensive Analysis of Deep Regression

By Stéphane Lathuilière, Pablo Mesejo, Xavier Alameda-Pineda, Radu Patrice Horaud. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018

330 citations recorded by the source corpus.

Deep learning revolutionized data science, and recently its popularity has grown exponentially, as did the amount of papers employing deep networks. Vision tasks, such as human pose estimation, did not escape from this trend. There is a large number of deep models, where small changes in the network architecture, or in the data pre-processing, together with the stochastic nature of the optimization procedures, produce notably different results, making extremely difficult to sift methods that significantly outperform others. This situation motivates the current study, in which we perform a systematic evaluation and statistical analysis of vanilla deep regression, i.e., convolutional neural networks with a linear regression top layer. This is the first comprehensive analysis of deep regression techniques. We perform experiments on four vision problems, and report confidence intervals for the median performance as well as the statistical significance of the results, if any. Surprisingly, the variability due to different data pre-processing procedures generally eclipses the variability due to modifications in the network architecture. Our results reinforce the hypothesis according to which, in general, a general-purpose network (e.g., VGG-16 or ResNet-50) adequately tuned can yield results close to the state-of-the-art without having to resort to more complex and ad-hoc regression models.

Subfields: Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/76cd24bc7ab988abef1ecd4c41d6cf65ba70c5fe

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

Energy-Based Models for Deep Probabilistic Regression (2020, background): “Common choices include the L(2) loss, `(ŷ, y) = ‖ŷ − y‖(2)2, the L(1) loss, `(ŷ, y) = ‖ŷ − y‖1, and their close relatives [21, 27].”
A Review on Deep Learning in UAV Remote Sensing (2021, background): “One approach [111] performed a comprehensive analysis of deep regression methods and pointed out that well-known fine-tuned networks, like VGG-16 [192] and ResNet-50 [75], can provide interesting results.”
Refine and Distill: Exploiting Cycle-Inconsistency and Knowledge Distillation for Unsupervised Monocular Depth Estimation (2019, methodology): “In all our experiments we use a batch size equal to 8 stereo image pairs and the Adam optimizer with learning rate set to 10−5 following the recommendations of [22].”
How to Train Your Energy-Based Model for Regression (2020, background): “Regression is a fundamental problem within computer vision with many additional applications [25, 42, 45, 54, 56], which all would benefit from such accurate predictions.”