Speridlabs · Research — NO. 01

ENEAS

Embedding-guided Neural Ensemble for Adaptive Segmentation

FIG. 01 — Demonstration reel1080 × 1080 · 46 s

ENEAS is a unified, text-promptable method for instance tracking and semantic discovery. Segmentation models — SAM 3 included — still lose a target that leaves the view, fragment it at extreme close-up, and mistake statues, paintings, or reflections for the real thing. ENEAS tells true instances from their doppelgangers.

arXiv Code Spaces

Javier del Pino^†Salvador RodríguezAlejandro GarabitoJavier ÁlvarezChema GarabitoSperidlabs · † project lead

02 — Method

Method

ENEAS works two ways from a single method: precise tracking and high-quality segmentation of a unique instance, and open-concept discovery of every instance a text query names. It takes a natural-language prompt and a set of frames — an ordered video or an unordered image collection — and returns the masks.

02.A

Instance Tracking

Name one instance and ENEAS follows it through the whole sequence — through occlusion and extreme changes of scale — holding onto the same object without drifting to look-alikes. It excels when the target leaves the frame and comes back: the moment it reappears, ENEAS picks it up again as the same instance.

Prompt “woman in the painting”848 × 480

Prompt “man”720 × 1280

02.B

Semantic Discovery

Name a category and ENEAS discovers and returns every instance of it, picking up new ones as they appear in the scene — and leaving out the doppelgangers: statues, paintings, or reflections that look like the target but are not.

Prompt “chair”720 × 1280

Prompt “people”848 × 480 · 0.5×

03 — Results

Results

A broad SA-V Test selection from SA-Co/VEval, split by prompt: prompts that refer to one specific object are assigned to instance tracking, prompts that may match multiple objects to semantic discovery. Both methods run under the official SAM 3 evaluator, identical settings. Best value per row in bold.

Table 1. SA-Co/VEval — Instance Tracking

Metric	ENEAS	SAM 3
HOTA (%)	26.70	26.51
DetA (%)	7.92	7.84
AssA (%)	90.77	90.18
LocA (%)	87.86	89.22
TETA (%)	17.86	16.65

Table 2. SA-Co/VEval — Semantic Discovery

Metric	ENEAS	SAM 3
HOTA (%)	9.23	9.19
DetA (%)	6.16	6.21
AssA (%)	14.28	13.85
LocA (%)	74.25	75.38
TETA (%)	9.53	9.10

04 — Comparisons

ENEAS vs SAM 3.

Same prompt, same footage, side by side.

ENEAS

SAM 3

Prompt“nearest guy with a gun”

ENEAS

SAM 3

Prompt“soldier outside the line”

05 — Citation

Cite this work

@article{delpino2026eneas,
  title  = {ENEAS: Embedding-guided Neural Ensemble
            for Adaptive Segmentation},
  author = {del Pino, Javier and Rodr\'iguez, Salvador
            and Garabito, Alejandro and \'Alvarez, Javier
            and Garabito, Chema},
  journal = {arXiv preprint arXiv:2609.03756},
  year   = {2026},
  url    = {https://arxiv.org/abs/2609.03756}
}