Regularizing face verification nets for pain intensity regression

Abstract

Limited labeled data are available for the research of estimating facial expression intensities. For instance, the ability to train deep networks for automated pain assessment is limited by small datasets with labels of patient-reported pain intensities. Fortunately, fine-tuning from a data-extensive pretrained domain, such as face verification, can alleviate this problem. In this paper, we propose a network that fine-tunes a state-of-the-art face verification network using a regularized regression loss and additional data with expression labels. In this way, the expression intensity regression task can benefit from the rich feature representations trained on a huge amount of data for face verification. The proposed regularized deep regressor is applied to estimate the pain expression intensity and verified on the widely-used UNBC-McMaster ShoulderPain dataset, achieving the state-of-the-art performance. A weighted evaluation metric is also proposed to address the imbalance issue of different pain intensities.

Forecasting Root-Zone Electrical Conductivity of Nutrient Solutions in Closed-Loop Soilless Cultures via a Recurrent Neural Network Using Environmental and Cultivation Information

“…In this study, mean square error (MSE) was used as a cost. Empirically, regressions based on ANNs usually use MSE instead of root-mean-square error (RMSE) as a cost for reducing computation (Esfe et al, 2016; Wang F. et al, 2017). The coefficient of determination ( R 2 ) was used for training and test accuracy.…”

Saliency Supervision: An Intuitive and Effective Approach for Pain Intensity Regression

“…Methods MAE↓ MSE↓ PCC↑ our proposed method 1 0.401 0.742 0.643 our proposed method 2 0.334 0.626 0.804 smooth L1 + L1 center loss [3] 0.456 0.804 0.651 OSVR-L1 [21] 1.025 N/A 0.600 OSVR-L2 [21] 0.810 N/A 0.601 RCR [13] N/A 1.54 0.65 Table 1. Performance of our proposed methods and related works on the UNBC-McMaster Shoulder-Pain dataset for the estimation of pain intensity.…”

“…Methods wMAE↓ wMSE↓ our proposed method 1 0.883 1.697 our proposed method 2 0.727 1.566 smooth L1 + L1 center loss + sampling [3] 0.991 1.720 OSVR-L1 [21] 1.309 2.758 OSVR-L2 [21] 1.299 2.719 Table 2. Performance of our network when evaluated using the weighted MAE and weighted MSE proposed by [3]. Methods proposed by us are applied with uniform class sampling technique.…”

“…Finally we finetune the network from the bottleneck layer supervised by pain intensity with regression loss, and get better result than the previous work [3].…”

“…Moreover, in most frames patients got zero pain intensity, making it hard to observe the pattern behind frames. Many methods [3] [4] have exploited deep neural networks trained in a data-extensive domain to alleviate the limited training data problem. For example, they finetune a well-trained face verification network with a regularized regression loss for pain intensity regression or valence-arousal estimation.…”