Adaloss: Adaptive Loss Function for Landmark Localization

Abstract

Landmark localization is a challenging problem in computer vision with a multitude of applications. Recent deep learning based methods have shown improved results by regressing likelihood maps instead of regressing the coordinates directly. However, setting the precision of these regression targets during the training is a cumbersome process since it creates a trade-off between trainability vs localization accuracy. Using precise targets introduces a significant sampling bias and hence makes the training more difficult, whereas using imprecise targets results in inaccurate landmark detectors. In this paper, we introduce "Adaloss", an objective function that adapts itself during the training by updating the target precision based on the training statistics. This approach does not require setting problem-specific parameters and shows improved stability in training and better localization accuracy during inference. We demonstrate the effectiveness of our proposed method in three different applications of landmark localization: 1) the challenging task of precisely detecting catheter tips in medical X-ray images, 2) localizing surgical instruments in endoscopic images, and 3) localizing facial features on in-the-wild images where we show state-of-the-art results on the 300-W benchmark dataset.

References in corpus (8)

• Adam: A Method for Stochastic Optimization
• ADADELTA: An Adaptive Learning Rate Method
• DeepPose: Human Pose Estimation via Deep Neural Networks
• Stacked Hourglass Networks for Human Pose Estimation
• An Empirical Study into Annotator Agreement, Ground Truth Estimation, and Algorithm Evaluation
• Style Aggregated Network for Facial Landmark Detection
• Unite the People: Closing the Loop Between 3D and 2D Human Representations
• Pose-Invariant Face Alignment with a Single CNN

Cited by in corpus (4)

• Descending through a Crowded Valley - Benchmarking Deep Learning Optimizers
• landmarker: a Toolkit for Anatomical Landmark Localization in 2D/3D Images
• Spine Landmark Localization with combining of Heatmap Regression and Direct Coordinate Regression
• Separable Batch Normalization for Robust Facial Landmark Localization with Cross-protocol Network Training