ProtoDCS: Towards Robust and Efficient Open-Set Test-Time Adaptation for Vision-Language Models

Name: ProtoDCS: Towards Robust and Efficient Open-Set Test-Time Adaptation for Vision-Language Models
Uploaded: 2026-10-03T00:47:24.541581+00:00
Duration: 1130 s
Description: ProtoDCS is a framework for adapting Vision-Language Models to new data during testing when the test set contains both known and unknown samples. It uses a four-stage process involving dual-threshold filtering, uncertainty-aware loss, and Gaussian Mixture Model verification to safely separate in-distribution (csID) from out-of-distribution (csOOD) data.

ProtoDCS is a framework for adapting Vision-Language Models to new data during testing when the test set contains both known and unknown samples. It uses a four-stage process involving dual-threshold filtering, uncertainty-aware loss, and Gaussian Mixture Model verification to safely separate in-distribution (csID) from out-of-distribution (csOOD) data.

Open the full episode