Delving into High-Quality Synthetic Face Occlusion Segmentation Datasets

This paper performs comprehensive analysis on datasets for occlusion-aware face segmentation, a task crucial for many downstream applications. Because collecting and annotating such data are time-consuming and labor-intensive, prior synthetic efforts often overlook naturalistic realism. We propose two occlusion generation techniques: Naturalistic Occlusion Generation (NatOcc), which produces high-quality, naturalistic synthetic occluded faces via color transfer, image harmonization, and super-resolution; and Random Occlusion Generation (RandOcc), a general synthetic method overlaying random shapes with textures and transparency. We empirically show both methods are effective and robust, including for unseen occlusions. To facilitate evaluation, we introduce two high-resolution real-world datasets with fine-grained masks: RealOcc (aligned, cropped) and RealOcc-Wild (in-the-wild) for robustness. We benchmark multiple segmentation models and provide analyses to guide future work. Code and datasets: https://github.com/kennyvoo/face-occlusion-generation.

Kenny T. R. Voo, Liming Jiang, Chen Change Loy