Amanda Adkins

PreSIST: Vision-Language-Informed Object Persistence Prediction in Open-World Scenes

This work proposes PreSIST (Predictive Scene-conditioned Instance Survival over Time), a method for predicting whether an observed object will remain in its last seen pose at arbitrary future times and develops two interchangeable variants: PreSIST-Lang, which estimates persistence priors using a VLM, and PreSIST-Vis,...