Knowledge-Preserved Model Tuning in Null-Space for Robust Spatio-Temporal Video Grounding
Researchers have developed a new method called Null-Space Tuning (NST) to improve spatio-temporal video grounding models. This technique addresses the issue of adapting pre-trained models to low-quality video inputs without disrupting their existing knowledge. NST achieves this by injecting learnable residuals into input features, which are selectively invisible to the frozen backbone of the model, thereby preserving pre-trained knowledge while rectifying degraded inputs. AI
IMPACT Enhances robustness of video analysis models to degraded input quality, potentially improving real-world applications.