Bıçakçı, Yunus Serhat · Oct 5, 2026
arXiv cs.AI: Vision Transformer Ensembles for Panoramic Street Segmentation
Markdown · JSON · Original source
Yunus Serhat Bıçakçı compares segmentation systems for describing street panoramas in the PalmCity challenge, where small datasets and unequal training costs make model selection difficult. In the first stage, nine systems were compared under approximately equal computation budgets. The leading approach averaged three Encoder only Mask with DINOv3 features, evaluated at three image scales with horizontal reflection. In the second stage, larger inputs produced the largest gains among the tested changes. The selected ensemble of four models scored 64.53% mean intersection over union on validation when trained on training images alone. Retrained on training and validation images, it scored 61.54% mean intersection over union and 71.93% mean F1 on the hidden test leaderboard. Bıçakçı reports that this ranked first among five entries in the 30 October 2026 snapshot. The validation score is not an independent check: validation images guided many choices, and the final models were trained on them. The paper says the hidden test score is the only independent measurement. It also reports negative results: Segment Anything voting, instance mask overlays, and ensembles with other architectures reduced the validation score. The results describe this challenge and its evaluation; how well they transfer beyond it remains open.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Oct 5, 2026 (2026-10-05).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T11:38:13.533Z. First observed: 2026-11-02T05:03:25.759Z. Retrieved: 2026-11-02T08:04:05.331Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The reported scores and leaderboard placements concern the PalmCity challenge; the evidence does not establish performance on additional independent street-panorama data.
Open question
Would the selected ensemble retain its advantage on additional independent street-panorama data?
Source-stated
The paper compares nine pretrained segmentation systems under approximately equal computation budgets, then evaluates leading candidates with three random seeds.
nine pretrained segmentation systems are compared using approximately equal computation budgets, and the two leading candidates are trained with three random seeds
Source-stated
The first-stage leading approach averages three Encoder only Mask Transformer models with DINOv3 features and scores 57.08% on the hidden test leaderboard snapshot of 5 October 2026.
An equal probability average of three Encoder only Mask Transformer models with DINOv3 features, evaluated at three image scales with horizontal reflection, reaches 60.95% mean intersection over union on the 84 image validation split and 57.08% on the hidden test leaderboard
Source-stated
The paper reports that training at larger input sizes gave the largest gains in the second stage.
Training at 768 by 1536 and 896 by 1792 pixels gives the largest gains.
Source-stated
The selected four-model ensemble scored 64.53% mean intersection over union on validation when trained on training images alone, and 61.54% mean intersection over union and 71.93% mean F1 on the hidden test leaderboard after retraining on training and validation images.
The selected ensemble of four models scores 64.53% on validation when trained on the training images alone. Retrained on the training and validation images, it receives 61.54% mean intersection over union and 71.93% mean F1 on the hidden test leaderboard
Source-stated
The paper reports that the selected ensemble ranked first among five entries in the 30 October 2026 snapshot.
first among five entries in the snapshot of 30 October 2026
Source-stated
The author cautions that validation scores are optimistic because validation images guided choices and were used to train final models; the hidden test score is the only independent measurement.
Because the validation images guided many choices and the final models were trained on them, the validation scores are optimistic and the test score is the only independent measurement.
Source-stated
The tested Segment Anything voting, instance mask overlays, and ensembles with other architectures reduced the validation score.
Segment Anything voting, instance mask overlays, and ensembles with other architectures all reduce the validation score.
Source and contributor context
Bıçakçı, Yunus Serhat · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Pre-training
A model's initial training on a large dataset, before further adaptation for particular uses.
Transformer architecture
The neural-network design behind most current language models; it weighs how each part of the input relates to the others.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: aa78903c-4d29-4011-8d01-863f9110acd9.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T11:38:13.533Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 6d66e594680ae63c037dfda9b5200c6e1a10f1a1ee45f53afa726906a1b881fe. Source text hash: 8d14ad375c4998663d706ea4760ea4546e6ff72c63b1ace996d769be97f01bcb.
Revision and source-check history
Version 1 · 2026-11-10T11:38:13.533Z
arXiv cs.AI: Vision Transformer Ensembles for Panoramic Street Segmentation
Yunus Serhat Bıçakçı compares segmentation systems for describing street panoramas in the PalmCity challenge, where small datasets and unequal training costs make model selection difficult. In the first stage, nine systems were compared under approximately equal computation budgets. The leading approach averaged three Encoder only Mask with DINOv3 features, evaluated at three image scales with horizontal reflection. In the second stage, larger inputs produced the largest gains among the tested changes. The selected ensemble of four models scored 64.53% mean intersection over union on validation when trained on training images alone. Retrained on training and validation images, it scored 61.54% mean intersection over union and 71.93% mean F1 on the hidden test leaderboard. Bıçakçı reports that this ranked first among five entries in the 30 October 2026 snapshot. The validation score is not an independent check: validation images guided many choices, and the final models were trained on them. The paper says the hidden test score is the only independent measurement. It also reports negative results: Segment Anything voting, instance mask overlays, and ensembles with other architectures reduced the validation score. The results describe this challenge and its evaluation; how well they transfer beyond it remains open.
Original source
cs.AI · Article
Source details
Original source date: Oct 5, 2026 (2026-10-05).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T11:38:13.533Z. First observed: 2026-11-02T05:03:25.759Z. Retrieved: 2026-11-02T08:04:05.331Z.
Limits
The reported scores and leaderboard placements concern the PalmCity challenge; the evidence does not establish performance on additional independent street-panorama data.
Open question
Would the selected ensemble retain its advantage on additional independent street-panorama data?
Source-stated
The paper compares nine pretrained segmentation systems under approximately equal computation budgets, then evaluates leading candidates with three random seeds.
nine pretrained segmentation systems are compared using approximately equal computation budgets, and the two leading candidates are trained with three random seeds
Source-stated
The first-stage leading approach averages three Encoder only Mask Transformer models with DINOv3 features and scores 57.08% on the hidden test leaderboard snapshot of 5 October 2026.
An equal probability average of three Encoder only Mask Transformer models with DINOv3 features, evaluated at three image scales with horizontal reflection, reaches 60.95% mean intersection over union on the 84 image validation split and 57.08% on the hidden test leaderboard
Source-stated
The paper reports that training at larger input sizes gave the largest gains in the second stage.
Training at 768 by 1536 and 896 by 1792 pixels gives the largest gains.
Source-stated
The selected four-model ensemble scored 64.53% mean intersection over union on validation when trained on training images alone, and 61.54% mean intersection over union and 71.93% mean F1 on the hidden test leaderboard after retraining on training and validation images.
The selected ensemble of four models scores 64.53% on validation when trained on the training images alone. Retrained on the training and validation images, it receives 61.54% mean intersection over union and 71.93% mean F1 on the hidden test leaderboard
Source-stated
The paper reports that the selected ensemble ranked first among five entries in the 30 October 2026 snapshot.
first among five entries in the snapshot of 30 October 2026
Source-stated
The author cautions that validation scores are optimistic because validation images guided choices and were used to train final models; the hidden test score is the only independent measurement.
Because the validation images guided many choices and the final models were trained on them, the validation scores are optimistic and the test score is the only independent measurement.
Source-stated
The tested Segment Anything voting, instance mask overlays, and ensembles with other architectures reduced the validation score.
Segment Anything voting, instance mask overlays, and ensembles with other architectures all reduce the validation score.
Source and contributor context
Bıçakçı, Yunus Serhat · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Pre-training
A model's initial training on a large dataset, before further adaptation for particular uses.
Transformer architecture
The neural-network design behind most current language models; it weighs how each part of the input relates to the others.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: aa78903c-4d29-4011-8d01-863f9110acd9.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T11:38:13.533Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 6d66e594680ae63c037dfda9b5200c6e1a10f1a1ee45f53afa726906a1b881fe. Source text hash: 8d14ad375c4998663d706ea4760ea4546e6ff72c63b1ace996d769be97f01bcb.
passed · 2026-11-10T11:38:13.533Z · anthropic/claude-opus-5.5
Publication: aa78903c-4d29-4011-8d01-863f9110acd9
attribution — pass: The reading correctly attributes the paper to Yunus Serhat Bıçakçı on arXiv. It keeps the leaderboard ranking and the validation caveat attributed to the author and paper rather than stating them as established facts. The 'arXiv cs.AI' label matches the listed publisher; the paper is cross-listed in cs.AI.
Authors:Yunus Serhat Bıçakçı
summary — pass: The summary faithfully covers both stages, the scores, the first-place ranking among five entries, the negative results and the author's caution about optimistic validation scores. It also opens by saying segmentation can describe street panoramas. Terms such as mean intersection over union, Encoder only Mask Transformer, DINOv3 and mean F1 are kept exactly, as the term-library tooltip approach requires. Omitting the release of code and models is minor. The question about transfer to new data follows from the paper's own caveat.
Semantic segmentation of street panoramas can support detailed descriptions of urban environments
date — pass: The original publication date of 2026-10-05 matches the v1 submission date. It is not the 30 October v2 revision or the November feed date.
[Submitted on 5 Oct 2026 (v1), last revised 30 Oct 2026 (this version, v2)]
support — pass: All claims are directly supported by verbatim source excerpts. The limitation and the question about transfer beyond the challenge are framed as open issues, not as asserted facts. Saying candidates are 'evaluated' with three seeds slightly paraphrases 'trained' but does not change the meaning.
the validation scores are optimistic and the test score is the only independent measurement
media — pass: The selected media is the verified arXiv abstract page for this paper, which is the direct original source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A model's initial training on a large dataset, before further adaptation for particular uses.
Definition referenceThe neural-network design behind most current language models; it weighs how each part of the input relates to the others.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA model's initial training on a large dataset, before further adaptation for particular uses.
Definition referenceThe neural-network design behind most current language models; it weighs how each part of the input relates to the others.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv