Mask Propagation With Confidence Based Filtering

via “mask propagation with confidence-based filtering”

Meta's foundation model for visual segmentation.

Unique: Implements confidence-based filtering on mask propagation to prevent error accumulation across frames, using model-estimated confidence scores to identify frames requiring re-prompting or manual correction. The filtering is applied post-prediction, enabling flexible threshold tuning without model retraining.

vs others: More practical than optical flow-based error detection because confidence scores are computed directly from the segmentation model, and more efficient than re-processing frames because filtering is applied selectively based on confidence rather than re-running inference on all frames.