Weakly-Supervised Attention and Relation Learning for Facial Action Unit Detection
share
Attention mechanism has recently attracted increasing attentions in the field of facial action unit (AU) detection. By finding the region of interest (ROI) of each AU with the attention mechanism, AU related local features can be captured. Most existing AU detection works design fixed attentions based on the prior knowledge, without considering the nonrigidity of AUs and relations among AUs. In this paper, we propose a novel end-to-end weakly-supervised attention and relation learning framework for AU detection, which has not been explored before. In particular, to select and extract AU related features, both channel-wise attentions and spatial attentions are learned with AU labels only. Moreover, pixel-level relations for AUs are learned to refine spatial attentions and extract more accurate local features. A multi-scale local region learning method is further proposed to adapt multi-scale AUs in different locations, which can facilitate the weakly-supervised attention and relation learning. Extensive experiments on BP4D and DISFA benchmarks demonstrate that our framework (i) outperforms the state-of-the-art methods for AU detection, and (ii) also achieves superior performance of AU intensity estimation with a simple extension.
READ FULL TEXT