Original title: ENEAS: Embedding-guided Neural Ensemble for Adaptive Segmentation
Javier del Pino, Salvador Rodríguez, Alejandro Garabito, Javier Álvarez +1 · September 2026
Computer vision models often fail when asked to find and outline objects using text prompts. Even modern foundation models hallucinate objects that left the screen, cut objects into pieces during close-ups, and get fooled by lookalikes like statues, paintings, or reflections. Researchers created ENEAS to solve this using a single approach that works two ways. First, for tracking, it builds on the geometrically sturdy SeC system by adding a text adapter and temporal memory. This memory holds onto a target across disappearances without drifting to distractors, keeping the item whole even when it fills the frame. Second, for discovering every item matching a query, a verification layer pairs fast visual embedding matching with a vision-language model. This checks only ambiguous cases to remove lookalikes while running quickly. ENEAS enables high-quality tracking and segmentation across video, broad libraries, and unordered datasets.
Imagine you're searching for your friend in a crowded mall, but there are mannequins dressed in identical outfits. At a glance, you might mistake a mannequin for your friend, wasting time or going to the wrong spot. In this paper, the system faces a similar challenge with video objects that look alike, like statues or reflections. Here, it uses text descriptions and a verification step to tell real objects from imposters, so it doesn't get fooled by visual similarities.
For creators building 3D models from video, mistaking a statue for a real person corrupts the entire model. This system helps by correctly identifying true objects, making video-based reconstruction more reliable.
The abstract does not provide quantitative results or direct comparisons to other methods. It does not specify computational efficiency or real-time performance capabilities. The work focuses on handling specific failure modes but does not address all possible visual distortions or data types.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
ENEAS: Embedding-guided Neural Ensemble for Adaptive Segmentation — Javier del Pino, Salvador Rodríguez, Alejandro Garabito et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.