ENEAS: Embedding-guided Neural Ensemble for Adaptive Segmentation
ENEAS is pitched as a text-promptable segmentation system meant to keep one object straight across video and reject lookalikes.
The paper says current text-promptable segmentation models can hallucinate objects over time, break objects into fragments, or mistake statues, paintings, and reflections for real target entities. ENEAS extends the SeC tracking architecture with text prompting and temporal memory so a target can disappear without the tracker drifting to a distractor. For broader discovery, it pairs fast visual embedding matches with VLM checks only on ambiguous candidates. The authors say the code and models are available. HF Daily Papers' note
The paper says current text-promptable segmentation models can hallucinate objects over time, break objects into fragments, or mistake statues, paintings, and reflections for real target entities. ENEAS extends the SeC tracking architecture with text prompting and temporal memory so a target can disappear without the tracker drifting to a distractor. For broader discovery, it pairs fast visual embedding matches with VLM checks only on ambiguous candidates. The authors say the code and models are available. HF Daily Papers' note
score 4