Event-SAMA: Language-Grounded Video Object Segmentation on Neuromorphic Event Streams

Event cameras provide microsecond temporal resolution, high dynamic range, and no motion blur, yet language-grounded video segmentation has been developed exclusively on dense RGB frames: no dataset and no model exist for referring video object segmentation (RVOS) on event streams. We address this gap with a dataset and a baseline model. First, we present **Event-SAMA-239K**, a synthetic-event conversion of the SAMA-239K grounded video-chat corpus and of all additional SAMA training data: 30,650 video clips and 392,162 images, converted with optical-flow, FILM, or saccade upsampling followed by a linear event generator, with all mask, referring-expression, and dialogue annotations preserved. Second, we present **Event-SAMA**, an adaptation of the SAMA-4B architecture that consumes events natively through a two-channel polarity patch embedding initialized from the pretrained RGB weights and a zero-initialized event adapter. Trained end-to-end on the converted corpus, Event-SAMA reaches $24.37$ J&F on MeViS `valid_u` and $22.33$ J&F on the converted subset of ReVOS `valid`, establishing the first event-native RVOS baseline (RGB SAMA-4B: $55.4$ on MeViS). A seed-controlled ablation with a measured noise floor of $0.47$ J&F identifies three design choices that matter: training on a diverse source mix instead of MeViS alone ($+4.99$), RGB-mean initialization of the event stem ($+4.33$), and a trainable SAM2 mask decoder ($+2.76$); the event adapter, partial ViT unfreezing, and event polarity are within noise at this budget. A comparison run with the identical recipe on an earlier, partially converted snapshot of the corpus scores $1.38$ J&F higher on MeViS and $0.73$ higher on ReVOS. Both are single runs that also differ in schedule length, so we draw no conclusion about corpus scale from this pair.