Dynamic Domain Generalization

Abstract

Both social group detection and group emotion recognition in images are growing fields of interest, but never before have they been combined. In this work we aim to detect emotional subgroups in images, which can be of great importance for crowd surveillance or event analysis. To this end, human annotators are instructed to label a set of 171 images, and their recognition strategies are analysed. Three main strategies for labeling images are identified, with each strategy assigning either 1) more weight to emotions (emotion-based fusion), 2) more weight to spatial structures (group-based fusion), or 3) equal weight to both (summation strategy). Based on these strategies, algorithms are developed to automatically recognize emotional subgroups. In particular, K-means and hierarchical clustering are used with location and emotion features derived from a fine-tuned VGG network. Additionally, we experiment with face size and gaze direction as extra input features. The best performance comes from hierarchical clustering with emotion, location and gaze direction as input.

Parameter Exchange for Robust Dynamic Domain Generalization

“…However, the growing number of parameters would increase the risk of overfitting. Some studies have been developed to solve this problem by squeezing the dynamic parameters into a low-dimensional space [44,47]. Sun et al [44] decomposes the network parameters into a dynamic component and a static component to reduce the dynamic parameter space and achieve dynamic domain generalization (DDG).…”

“…Most existing arts aim to obtain a robust static model by projecting multiple source domains into a common distribution, namely domain-invariant space [31,35,42,56], where the domain-specific information is neglected, resulting in an adaptation failure to the unknown target domains (see Figure 1(a)). Fortunately, the limitation of static models has been alleviated by Dynamic Domain Generalization (DDG) [44], which develops a dynamic network to achieve training-free adaptation on the unknown target domains, as shown in Figure 1(b). By decoupling the network parameters into a static component and a dynamic component, DDG allows the static component to learn domain-invariant features shared across instances, while the dynamic component adapts to each instance on the fly, learning domain-specific features.…”

“…• Based on the existing art [44], we propose a simple yet effective Parameter Exchange method that aims to disentangle the domaininvariant and -specific features more thoroughly by perturbing the combination between the static and dynamic components. • Specifically, we develop two different PE manners: cross-instance PE that exchanges the parameters of dynamic component across different instances, and cross-kernel PE that reassigns them across different kernels.…”

“…Also, to facilitate the feature disentanglement, the model is updated using the gradients from both the perturbed and non-perturbed feed-forward jointly. • Extensive experimental results show that PE is a plug-andplay method that can be easily applied to existing dynamic networks, such as DDG [44], DRT [32], and ODConv [24], to significantly improve their generalization performance on different DG benchmarks. In particular, our method integrated with ODConv achieves state-of-the-art performance on multiple DG benchmarks, indicating the effectiveness of PE in advancing the dynamic domain generalization.…”

Attention Diversification for Domain Generalization

“…However, the domain shift is usually agnostic in real-world scenarios since the target data is not available for training. This issue inspires the research area of domain generalization (DG) [1,22,27,28,30,34,41,43,45,47,51,52,54,74,75,[78][79][80], which is aimed to make models trained on seen domains achieve accurate predictions on unseen domainss, i.e., the conditional distribution P (Y |X) is robust with shifted marginal distribution P (X). Canonical DG focuses on learning a domaininvariant feature distribution P (F (X)) across domains for the robustness of conditional distribution P (Y |F (X)).…”