Unsupervised domain adaptation (UDA) transfers knowledge from labeled datasets (sources) to a new unlabeled dataset (target). While image-based UDA is well-studied, video-based UDA is less explored due to the complexity of adapting diverse video features and modeling temporal associations. Current methods use optical flow for motion cues but are computationally intensive and domain-specific.
We propose an adversarial domain adaptation approach for video semantic segmentation that aligns temporally associated pixels across source and target frames without optical flow. Our Perceptual Consistency Matching (PCM) strategy leverages perceptual similarity to identify correlated pixels across frames, assuming they correspond to the same class. This enhances prediction accuracy by enforcing consistency within and across domains during training. Extensive experiments on public datasets demonstrate our method's superiority over state-of-the-art UDA methods, offering improved performance and faster inference times.
Негізгі бет Ойындар Video Domain Adaptation for Semantic Segmentation using Perceptual Consistency Matching
Пікірлер