Answers are grounded in this paper and cite their sources.
This work focuses on the segmentation of fixed overlay logos in videos. Although state-of-the-art video semantic segmentation models achieve good results, they often rely on complex architectures with high memory and computational costs. These constraints are even more critical in video processing, where models must handle temporal consistency and efficiently process multiple frames. However, applications such as discriminating overlay content from scene elements have specific constraints and exploitable priors that make generalist models less suitable. We present a hybrid approach that addresses this task by combining the adaptability of deep learning with handcrafted spatio-temporal features. This hybrid architecture outperforms traditional models that process frames directly or rely solely on isolated cues, while maintaining competitive inference times. Extensive experiments confirm its effectiveness. Our code is available at https://gitlab.insa-rennes.fr/qmonnier/hybrid-approach-for-logo-segmentation-in-videos/.
No takes yet. Share an insight, caveat, or question.
Monnier et al. (2025) studied this question.