Light-Weight Distilled HRNet for Facial Landmark Detection

A light-weight facial landmark detection model is proposed in this paper (we named it “LDHRNet”), which can be trained in an end-to-end fashion and could perform precise facial landmark detection in various conditions including those with large pose, exaggerated expression, non-uniform lighting and occlusions. Firstly, in order to deal with these challenging cases above, a light-weight HRNet (LHRNet) structure is proposed as the backbone while the bottleneck block is used to replace the standard residual block in the original HRNet and the group convolution is used to replace the standard convolution in the original HRNet. Then in order to prevent the accuracy loss by the coordinates quantization, we use function named dual soft argmax (DSA) to map the heatmap response to final coordinates. And then we proposed Similarity-FeatureMap knowledge distillation model which guides the training of a student network such that input pairs that produce similar (dissimilar) feature maps in the pre-trained teacher network produce similar (dissimilar) feature maps in the student network. Finally, we combine the distillation loss and NME loss to train our model. The best result 79.10% for AUC is achieved on the validation set.