ENEAS: Embedding-guided Neural Ensemble for Adaptive Segmentation
Javier del Pino, Salvador Rodríguez, Alejandro Garabito, Javier Álvarez, Chema Garabito
cs.CV · cs.AI
Quick verdict
ENEAS: text-prompted tracking plus a VLM verification cascade cuts SAM 3's false positives (F1 19.5→82.8 on one capture); promising, single-dataset evidence.
Summary
ENEAS is a text-promptable segmentation system with two modes: instance tracking and semantic discovery. For tracking, the authors add a text-driven initialization (Florence-2 grounding) to the SeC memory tracker, claiming robustness to target absence and extreme scale change. For discovery, a cascade scores Florence-2 region proposals with SigLIP 2 sigmoid embeddings (prompt ensembles), accepts or discards clear cases, and routes only an uncertainty interval to a Qwen3-VL judge (2B default, 4B optional) with neighbor masking, before SAM 2 produces masks. Quantitative evidence centers on Church Statues, a self-collected capture of hyper-realistic sculptures annotated by the authors, where SAM 3 scores F1 19.5% (precision 11.1%) versus 82.8% for ENEAS-2B and 87.6% for ENEAS-4B. Ablations (Table 5) show selective VLM consultation beats always-on judging on both latency (3.29 vs 5.05 s/frame on an L4) and F1 (82.8 vs 79.1). On a SA-Co/VEval subsample, ENEAS marginally improves tracking HOTA over SAM 3 (26.70 vs 26.51) mainly via association, with unexplained, very low DetA (about 6-8%). Main limitations: evidence rests on one small author-annotated dataset plus qualitative sequences; no SAM 3-plus-verifier baseline; no variance or significance testing; discovery lacks cross-frame identity; not real-time; recall is capped by region proposals.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 2
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Adding a text-driven initialization (Florence-2 grounding) to the SeC memory tracker yields a text-promptable tracker that reports target absence instead of drifting and preserves spatial integrity under extreme close-ups (Sections 3.1, 4.4).
- A cascade of Florence-2 proposals, SigLIP 2 sigmoid scoring with prompt ensembles, and a Qwen3-VL judge consulted only inside an uncertainty interval removes ontological false positives (statues, paintings) that visual-only models cannot distinguish (Section 3.2).
- On Church Statues, this raises F1 from 19.5% (SAM 3, precision 11.1%) to 82.8% with a 2B judge and 87.6% with a 4B judge at precision above 94% (Table 1).
- Selective VLM consultation is both faster (3.29 vs 5.05 s/frame) and more accurate (F1 82.8 vs 79.1) than always consulting the judge, evidencing over-reasoning when unambiguous crops are adjudicated by the VLM (Table 5).
- On a SA-Co/VEval subsample, ENEAS matches or slightly improves SAM 3's tracking profile (HOTA 26.70 vs 26.51; TETA 17.86 vs 16.65), mainly through association quality (Tables 2-3).
Strengths
Problem decomposition maps cleanly onto design
Section 1 names three concrete SAM 3 failure modes (temporal hallucination, spatial fragmentation, ontological misclassification) and Sections 3.1-3.2 address each with a distinct mechanism: SeC temporal memory for identity, single-anchor propagation for spatial integrity, conditional VLM judging for ontology.
Honest and informative ablations
Table 4 quantifies every component (RPN F1 18.6% → +embeddings strict 78.6% → +VLM 82.8%), and Table 5 documents a genuine 'over-reasoning' effect: VLM-Only is both slower (5.05 s) and less accurate (F1 79.1%) than the selective cascade (3.29 s, 82.8%), with activation rates (16.2%-100%) reported per configuration.
Realistic cost accounting
Well-motivated adversarial testbed and candid limitations
Weaknesses
Quantitative evidence rests on one small author-annotated dataset
Decisive baseline is missing
Public-benchmark gains are near-ties on an unspecified subsample with an anomalous metric
No statistical treatment anywhere
Tracking-mode claims are qualitative-only
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
ENEAS (SperidLabs, arXiv 2609.03756, September 2026) is a technical report presenting a unified text-promptable segmentation method with two modes sharing one component set: (1) instance tracking — text-driven initialization of the SeC memory tracker [85] via Florence-2 grounding, claimed to report target absence and survive extreme close-ups; and (2) semantic discovery — a per-frame cascade of Florence-2 proposals, SigLIP 2 sigmoid scoring with prompt ensembles, a conditional Qwen3-VL judge, and SAM 2 masks. The central thesis is that SAM 3-class failures stem from 'strong perception coupled with weak verification,' and that routing only an uncertainty interval (robust preset 0.10-0.90) to a small VLM judge removes ontological false positives at low latency. The abstract's claims are internally consistent with the presented tables, but every headline number is generated in-house.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is ENEAS and what problem does it solve?
ENEAS is a text-promptable segmentation method from SperidLabs that takes a text prompt plus video frames (ordered or unordered) and returns object masks. It targets three failures of models like SAM 3: hallucinating targets after they leave the view, fragmenting an object during extreme close-ups, and segmenting look-alikes (statues, paintings, reflections) as the real object. It works in two modes: tracking one specific named instance over time, and discovering every instance of a named category in each frame.
How much better is ENEAS than SAM 3?
On the paper's own Church Statues capture (hyper-realistic statues next to real visitors, prompt 'person'), SAM 3 scores F1 19.5% with 11.1% precision, while ENEAS with a 2B judge scores 82.8% F1 and with a 4B judge 87.6%. On the public SA-Co/VEval benchmark subsample the two systems are nearly tied (tracking HOTA 26.70 vs 26.51), so the large advantage is specific to the ontologically ambiguous scenario.
Is ENEAS's code and data available?
The paper states that code and models are available at github.com/speridlabs/eneas. However, the three custom evaluation sequences (Church Statues, Moving Boxes, Blue Painting) — including the dataset behind all headline numbers — are not stated to be publicly released, which limits independent verification.
Is ENEAS real-time?
No. Depending on configuration it takes about 1.0 to 3.3 seconds per frame with the 2B judge on a single NVIDIA L4 GPU, and about 5 seconds with the 4B judge. The authors position it as an offline preprocessing tool for 3D reconstruction pipelines, not a real-time system.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 2-5 carry the qualitative evidence for the tracking and ontological-robustness claims (Sections 4.4-4.6); the images themselves could not be assessed from the extracted text, so all figure-based judgments rely on the captions.
- The text otherwise appears complete: full abstract, all six tables (values legible), and references [01]-[90] are present; no truncation detected.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.