Shaya, Mousa · Nov 16, 2026
Mousa Shaya asks how much SAR vehicle recognition depends on the vehicle
Markdown · JSON · Original source
Mousa Shaya’s paper tests whether SAR vehicle-recognition scores reflect the vehicle or shortcuts in the image. On MSTAR, a modern network still reaches 73.3% accuracy after the central 96x96 pixels are zeroed out; on ATRNet-STAR, accuracy is 63.3% with the vehicle removed. But under ATRNet-STAR’s official scene-change split, accuracy falls to 22.8% on full chips and rises to 30.8% when only the vehicle is shown. Shaya also reports that performance depends on the gap between training and test depression angles. In a setting where most classes have training images at the test angle and a few do not, the missing classes fare worse at a 45-degree gap than when no class has training images at that angle. The results challenge what a high score alone can establish about recognition: background cues and angle coverage affect performance, and the tested does not solve scene change. Shaya recommends reporting background-only and target-only controls, alongside the angle gap, with accuracy figures.
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: Nov 16, 2026 (2026-11-16).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-18T12:34:59.172Z. First observed: 2026-11-18T05:16:57.840Z. Retrieved: 2026-11-18T12:33:38.796Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The supplied evidence is the paper’s abstract; it does not provide the experimental setup and analysis needed to assess the reported results in detail. The abstract reports results on MSTAR and ATRNet-STAR; it does not establish how well they generalize to other datasets or operational settings.
Open question
What explains why classes missing training images at the test angle do worse when other classes have them than when no class does?
Source-stated
Mousa Shaya reports that on MSTAR, a modern network retains substantial accuracy when the central image region is zeroed out.
with the central 96x96 pixels zeroed out, accuracy is still 73.3% (chance 10%).
Source-stated
Shaya reports background-only performance on ATRNet-STAR, but under its official scene-change split, showing only the vehicle yields higher accuracy than using the full image.
Under the official scene-change split, the background stops helping and starts hurting. Accuracy is 22.8% on full chips and 30.8% when only the vehicle is shown.
Source-stated
Shaya reports that accuracy declines as the gap to the nearest training depression angle increases in the tested settings.
Accuracy is set mainly by the gap to the nearest training angle: 57% at 15 degrees, 39% at 30 degrees and 26% at 45 degrees.
Source-stated
When some classes lack training images at the test angle but other classes have them, Shaya reports lower accuracy for the missing classes than when no class has training images at that angle.
the missing classes do worse than when no class has them: 9.0% instead of 22.4% at a 45-degree gap.
Source-stated
Shaya recommends reporting background-only and target-only controls, as well as the angle gap, alongside accuracy.
We recommend reporting background-only and target-only controls, and the angle gap, next to any accuracy figure.
Source and contributor context
Shaya, Mousa · 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.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Foundation model
A large model trained on broad data that can be adapted to many different tasks.
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: 4f75312d-3938-4cf7-8614-f4df143211f7.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-18T12:34:59.172Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: b709f9e0896c356041bc82dad5f506cf6db0c03271682e109be2f26ad0e20c1f. Source text hash: 55b8ae4d2c8ace9483f210b001b547c982901133719bbf6ab27a4e444206941a.
Revision and source-check history
Version 1 · 2026-11-18T12:34:59.172Z
Mousa Shaya asks how much SAR vehicle recognition depends on the vehicle
Mousa Shaya’s paper tests whether SAR vehicle-recognition scores reflect the vehicle or shortcuts in the image. On MSTAR, a modern network still reaches 73.3% accuracy after the central 96x96 pixels are zeroed out; on ATRNet-STAR, accuracy is 63.3% with the vehicle removed. But under ATRNet-STAR’s official scene-change split, accuracy falls to 22.8% on full chips and rises to 30.8% when only the vehicle is shown. Shaya also reports that performance depends on the gap between training and test depression angles. In a setting where most classes have training images at the test angle and a few do not, the missing classes fare worse at a 45-degree gap than when no class has training images at that angle. The results challenge what a high score alone can establish about recognition: background cues and angle coverage affect performance, and the tested does not solve scene change. Shaya recommends reporting background-only and target-only controls, alongside the angle gap, with accuracy figures.
Original source
cs.AI · Article
Source details
Original source date: Nov 16, 2026 (2026-11-16).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-18T12:34:59.172Z. First observed: 2026-11-18T05:16:57.840Z. Retrieved: 2026-11-18T12:33:38.796Z.
Limits
The supplied evidence is the paper’s abstract; it does not provide the experimental setup and analysis needed to assess the reported results in detail. The abstract reports results on MSTAR and ATRNet-STAR; it does not establish how well they generalize to other datasets or operational settings.
Open question
What explains why classes missing training images at the test angle do worse when other classes have them than when no class does?
Source-stated
Mousa Shaya reports that on MSTAR, a modern network retains substantial accuracy when the central image region is zeroed out.
with the central 96x96 pixels zeroed out, accuracy is still 73.3% (chance 10%).
Source-stated
Shaya reports background-only performance on ATRNet-STAR, but under its official scene-change split, showing only the vehicle yields higher accuracy than using the full image.
Under the official scene-change split, the background stops helping and starts hurting. Accuracy is 22.8% on full chips and 30.8% when only the vehicle is shown.
Source-stated
Shaya reports that accuracy declines as the gap to the nearest training depression angle increases in the tested settings.
Accuracy is set mainly by the gap to the nearest training angle: 57% at 15 degrees, 39% at 30 degrees and 26% at 45 degrees.
Source-stated
When some classes lack training images at the test angle but other classes have them, Shaya reports lower accuracy for the missing classes than when no class has training images at that angle.
the missing classes do worse than when no class has them: 9.0% instead of 22.4% at a 45-degree gap.
Source-stated
Shaya recommends reporting background-only and target-only controls, as well as the angle gap, alongside accuracy.
We recommend reporting background-only and target-only controls, and the angle gap, next to any accuracy figure.
Source and contributor context
Shaya, Mousa · 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.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Foundation model
A large model trained on broad data that can be adapted to many different tasks.
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: 4f75312d-3938-4cf7-8614-f4df143211f7.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-18T12:34:59.172Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: b709f9e0896c356041bc82dad5f506cf6db0c03271682e109be2f26ad0e20c1f. Source text hash: 55b8ae4d2c8ace9483f210b001b547c982901133719bbf6ab27a4e444206941a.
uncertain · 2026-11-18T12:34:59.172Z · anthropic/claude-opus-5.5
Publication: 4f75312d-3938-4cf7-8614-f4df143211f7
attribution — pass: The paper lists Mousa Shaya as its sole author. The draft consistently attributes the findings and the recommendation to Shaya rather than presenting them as established fact.
Authors:Mousa Shaya
summary — uncertain: The numbers and the paper's structure are reported faithfully. However, a general reader would struggle with central terms that the draft never explains. 'SAR' (synthetic aperture radar) is the subject of the title and summary but is never expanded. 'Chips', 'depression angles', 'scene-change split' and 'foundation model' are also unexplained, so the core finding is hard to follow without specialist background. Smallest repair: briefly define SAR and depression angle in plain language, or add tooltips for them. A minor gap: the summary does not note that the 73.3% MSTAR figure reproduces a known result.
we first reproduce the known MSTAR result: with the central 96x96 pixels zeroed out, accuracy is still 73.3% (chance 10%).
date — pass: originalPublished is 2026-11-16, which matches the v1 submission date on arXiv. The feed date of 2026-11-18 is correctly not used.
[Submitted on 16 Nov 2026]
support — pass: Each claim excerpt matches the abstract. The framing that the results challenge what a high benchmark score alone can establish follows from the paper's controls and its recommendation. The question about why missing classes do worse is open and grounded in the source; the abstract reports only that most of their images are assigned to known classes. The limitations are appropriately scoped.
Most of their images are assigned to known classes.
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 defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA large model trained on broad data that can be adapted to many different tasks.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA large model trained on broad data that can be adapted to many different tasks.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv