SteerSeg: Attention Steering for Reasoning Video Segmentation

Researchers introduced SteerSeg, a lightweight method for text-guided video segmentation.

It steers attention in frozen vision-language models using learnable soft prompts and Chain-of-Thought guidance.

The model turns attention maps into segmentation cues and ranks candidate tracklets for better grounding.

SteerSeg improves accuracy and generalizes beyond its training data without retraining large models.