CapMap-MS-TTA: 3rd Place Solution for the MUMU Track of the 8th LSVOS Challenge at ECCV 2026
cs.CV · cs.AI
Quick verdict
Training-free Florence-2-base system (lexicon tagging + multi-scale flip TTA) takes 3rd in the MUMU challenge with modest, honestly ablated gains (15.16 to 16.48); no code, metrics underspecified.
Summary
This technical report presents CapMap-MS-TTA, the 3rd-place entry in the MUMU track of the 8th LSVOS Challenge (ECCV 2026), where one model under 0.5B parameters and 8GB memory must jointly perform hierarchical image tagging (Task A), open-vocabulary detection (Task B), and English captioning (Task C). The system is training-free, built on Florence-2-base (~231M parameters, ~0.6GB peak memory). Task C uses the native <DETAILED_CAPTION> prompt with whitespace normalization, a 300-character cap, and greedy truncation to 30 evaluator tokens, decoded with beam search 3 at batch size 1 for reproducibility. Task A maps the same caption into official vocabularies via a hand-curated whole-phrase lexicon, image-statistics quality priors (Laplacian variance, brightness), add-only expand-hints fusion, and indoor/outdoor exclusivity cleanup. Task B runs <OD> at scales 0.75/1.0/1.25 plus horizontal flip, rescales boxes to original coordinates, and applies label-aware NMS (IoU 0.5, max 300 boxes). On the 1,038-image public test split, the reproduced baseline (Final 15.1568) improves to 16.4815 (A 21.15 to 26.49, B 4.77 to 5.45, C unchanged at 44.05), attributed to Task A recall and denser detections (9.93 to 11.44 boxes/image). Limitations: hand-curated lexicon coverage, persistent empty detections, aggressive 30-token truncation; no significance analysis, undefined Final aggregation, and no code release.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 4
● low 2
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single training-free Florence-2-base checkpoint (~231M parameters, ~0.6GB peak memory) satisfies the MUMU 0.5B/8GB constraints while serving tagging, detection, and captioning.
- Caption-driven lexicon mapping with add-only expand-hints lifts Task A from 21.15 to 26.49, attributed almost entirely to recall gains.
- Multi-scale plus horizontal-flip OD TTA with label-aware NMS raises Task B from 4.77 to 5.45 and Final from 15.1568 to 16.4815, ranking 3rd on the final MUMU leaderboard.
- Task C remains constant at 44.05 across all ablation rows because its pathway is never modified, isolating the A/B contributions.
Strengths
Clean component isolation in ablations
Task C is held exactly at 44.05 across all Table 1 rows by never touching its pathway, so Task A/B gains (+5.34 and +0.68) can be attributed unambiguously; the authors state this isolation explicitly in Section 4.
Honest negative results and bug disclosure
Section 4 reports that the fine-tuned Florence-2-base-ft collapsed to ~8 Final, that bs>1 breaks caption reproducibility, and that early substring matching produced hundreds of false scene tags (ski/skin, mall/small) with the whole-phrase fix described.
Offline evidence corroborates leaderboard deltas
Quantified constraint compliance and determinism care
Weaknesses
Task B scores are very low and the metric is undefined
No statistical rigor behind 4-decimal improvements
Central assets unreleased despite reproducibility claim
Logically flawed precision claim for Task A
No competitive context
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 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 CapMap-MS-TTA?
CapMap-MS-TTA is a training-free challenge system for the MUMU track of the 8th LSVOS Challenge. It uses Microsoft's Florence-2-base model (~231M parameters) as a single shared model for image tagging, open-vocabulary object detection, and English captioning, adding caption-keyword lexicon mapping for tagging and multi-scale plus horizontal-flip test-time augmentation with NMS for detection.
What score did CapMap-MS-TTA achieve in the LSVOS MUMU challenge?
It scored Final 16.4815 (Task A tagging 26.49, Task B detection 5.45, Task C captioning 44.05) on the 1,038-image public MUMU test set, ranking 3rd on the final leaderboard. The reproduced Florence-2 baseline scored 15.1568, so the pipeline added about 1.32 points without any training.
Does CapMap-MS-TTA require fine-tuning?
No. The final submission uses the stock Florence-2-base checkpoint without fine-tuning; all gains come from prompt routing, deterministic post-processing (truncation, lexicon mapping, NMS), and test-time augmentation. The authors note their fine-tuned checkpoint attempt actually performed much worse (~8 Final).
Why are captions truncated to 30 tokens in the MUMU task?
The challenge evaluation enforces a 30-token caption limit, while Florence-2's detailed captions are typically 50-80 words. The pipeline normalizes whitespace, caps at 300 characters, and greedily drops trailing tokens to fit, keeping the leading subject and setting and sacrificing trailing attribute details.
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.
- Figure 1 is only available as a caption in the extracted text; the diagram itself was not analyzable.
- No supplementary material or appendix is present in the extraction, so additional implementation details may exist.
- The Final aggregation formula and Task B metric are absent from the text, limiting verification of headline numbers rather than the review's judgment of them.
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.