Adaptive robust loss for landmark detection

Abstract

Using a loss function with appropriate properties to supervise the training of the neural network can substantially improve its accuracy and speed. A majority of loss functions applied to landmark detection, including Smooth ℓ1 loss, Wing loss and Adaptive Wing loss, utilize a form of piece-wise function to enhance the loss value of small errors. Still, some issues arise with these functions, such as non-differentiable, sensitivity to the outliers and additional computation to guarantee the continuity of the functions. In this paper, we propose a unified loss function called Adaptive Robust loss (ARobust loss) that is tailored for two landmark detection paradigms, i.e., coordinate regression and heatmap regression. In addition, we demonstrate the robustness, smoothness and continuity of the proposed loss function towards the two paradigms analytically. Last but not least, we validate the preeminence and accuracy of the adaptive robust loss on several landmark detection datasets by experiments. The best NME(%) results under the HRNet-W18 reached 4.52 (WFLW-Test) and 3.34 (300W-Full), respectively. In the standard task of face alignment and human pose estimation, the performance of our approach is on par with the SOTA losses. And the advantage of our proposed loss is more obvious in some challenging scenarios.