Hasan, S Mahmudul, Peng, Lu · Oct 26, 2026
A benchmark tests corruption robustness and deployment quantization in crowd counting
Markdown · JSON · Original source
Authors S Mahmudul Hasan and Lu Peng introduce CrowdCC, a that tests how image corruption and model affect crowd-counting models together. Quantization reduces the precision used in a model’s calculations to meet device limits. The benchmark evaluates 26 models across four datasets, using 23 corruptions at three severity levels, and measures throughput and energy on Jetson Orin NX configurations. The authors report that capture-chain corruptions cause the largest performance losses across the evaluated models. They also find that strong performance on clean inputs does not fully predict performance on corrupted inputs. INT8 changes corruption robustness more than FP16, but its effects have no consistent direction, and it does not consistently improve or energy use. The results suggest that choosing a model for deployment may require weighing its performance under relevant image degradation against measured device costs, rather than relying on clean-input accuracy or precision alone. The paper reports that INT8 efficiency is strongly associated with reformat overhead in the compiled engine; its abstract does not establish how broadly these findings apply beyond the tested models, datasets, and device configurations.
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 26, 2026 (2026-10-26).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-06T00:02:00.715Z. First observed: 2026-10-28T05:04:37.466Z. Retrieved: 2026-10-29T22:46:52.428Z.
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 abstract describes evaluation on four datasets, 26 models, and corresponding Jetson Orin NX configurations; it does not establish how well the findings generalize beyond those evaluations. The abstract reports associations between INT8 inference efficiency and reformat overhead, but does not establish that this relationship applies to other compiled engines or hardware.
Open question
Do the same trade-offs hold for crowd-counting models and deployment hardware beyond those evaluated?
Source-stated
Hasan and Peng introduce CrowdCC to evaluate image corruption and deployment quantization together for crowd-counting models.
We introduce CrowdCC, a controlled corruption benchmark that jointly evaluates these effects
Source-stated
The benchmark evaluates 23 deployment-motivated corruptions at three severity levels, across four datasets and 26 models, and includes throughput and energy measurements from Jetson Orin NX configurations.
using 23 deployment-motivated corruptions at three severity levels, four crowd-counting datasets, and 26 models
Source-stated
The authors report that capture-chain corruptions cause the largest degradation across the evaluated models, and that clean accuracy does not fully predict corrupted performance.
capture-chain corruptions consistently produce the largest degradation across the evaluated models
Source-stated
The authors report that INT8 changes corruption robustness substantially more than FP16, without a systematic direction, and does not consistently improve latency or energy.
INT8 alters corruption robustness substantially more than FP16, with no systematic direction, and does not consistently improve latency or energy.
Source and contributor context
Hasan, S Mahmudul, Peng, Lu · 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.
Quantization
Storing a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
Latency
The delay between sending a request and receiving the result.
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: 6f5dff21-3a01-400b-8958-afe6d78278f0.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-06T00:02:00.715Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: c831380c26cc68c0fcf5c4d23aa45df8ee90dc6fc655cf2616efb8a36e7c328b. Source text hash: 3658db128bb631b1d8458c001774ebb24ba5fb3722f5b591c05d6d1307d8d661.
Revision and source-check history
Version 1 · 2026-11-06T00:02:00.715Z
A benchmark tests corruption robustness and deployment quantization in crowd counting
Authors S Mahmudul Hasan and Lu Peng introduce CrowdCC, a that tests how image corruption and model affect crowd-counting models together. Quantization reduces the precision used in a model’s calculations to meet device limits. The benchmark evaluates 26 models across four datasets, using 23 corruptions at three severity levels, and measures throughput and energy on Jetson Orin NX configurations. The authors report that capture-chain corruptions cause the largest performance losses across the evaluated models. They also find that strong performance on clean inputs does not fully predict performance on corrupted inputs. INT8 changes corruption robustness more than FP16, but its effects have no consistent direction, and it does not consistently improve or energy use. The results suggest that choosing a model for deployment may require weighing its performance under relevant image degradation against measured device costs, rather than relying on clean-input accuracy or precision alone. The paper reports that INT8 efficiency is strongly associated with reformat overhead in the compiled engine; its abstract does not establish how broadly these findings apply beyond the tested models, datasets, and device configurations.
Original source
cs.AI · Article
Source details
Original source date: Oct 26, 2026 (2026-10-26).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-06T00:02:00.715Z. First observed: 2026-10-28T05:04:37.466Z. Retrieved: 2026-10-29T22:46:52.428Z.
Limits
The abstract describes evaluation on four datasets, 26 models, and corresponding Jetson Orin NX configurations; it does not establish how well the findings generalize beyond those evaluations. The abstract reports associations between INT8 inference efficiency and reformat overhead, but does not establish that this relationship applies to other compiled engines or hardware.
Open question
Do the same trade-offs hold for crowd-counting models and deployment hardware beyond those evaluated?
Source-stated
Hasan and Peng introduce CrowdCC to evaluate image corruption and deployment quantization together for crowd-counting models.
We introduce CrowdCC, a controlled corruption benchmark that jointly evaluates these effects
Source-stated
The benchmark evaluates 23 deployment-motivated corruptions at three severity levels, across four datasets and 26 models, and includes throughput and energy measurements from Jetson Orin NX configurations.
using 23 deployment-motivated corruptions at three severity levels, four crowd-counting datasets, and 26 models
Source-stated
The authors report that capture-chain corruptions cause the largest degradation across the evaluated models, and that clean accuracy does not fully predict corrupted performance.
capture-chain corruptions consistently produce the largest degradation across the evaluated models
Source-stated
The authors report that INT8 changes corruption robustness substantially more than FP16, without a systematic direction, and does not consistently improve latency or energy.
INT8 alters corruption robustness substantially more than FP16, with no systematic direction, and does not consistently improve latency or energy.
Source and contributor context
Hasan, S Mahmudul, Peng, Lu · 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.
Quantization
Storing a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
Latency
The delay between sending a request and receiving the result.
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: 6f5dff21-3a01-400b-8958-afe6d78278f0.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-06T00:02:00.715Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: c831380c26cc68c0fcf5c4d23aa45df8ee90dc6fc655cf2616efb8a36e7c328b. Source text hash: 3658db128bb631b1d8458c001774ebb24ba5fb3722f5b591c05d6d1307d8d661.
uncertain · 2026-11-06T00:02:00.715Z · anthropic/claude-opus-5.5
Publication: 6f5dff21-3a01-400b-8958-afe6d78278f0
attribution — pass: The draft correctly attributes CrowdCC to S Mahmudul Hasan and Lu Peng. It frames the findings as the authors' reports, matching the arXiv abstract.
Authors:S Mahmudul Hasan, Lu Peng
summary — uncertain: The summary is faithful, but central terms are left unexplained for a general reader: INT8, FP16, capture-chain corruptions, Jetson Orin NX, throughput, reformat overhead and compiled engine. The main INT8-versus-FP16 finding is hard to follow without knowing that these are lower-precision number formats. It also omits that the models run on edge devices, which explains why quantization matters. Smallest repair: add one sentence on edge deployment and define INT8/FP16 and capture-chain corruption, or add term-library tooltips for them. The softening of the source's 'must consider' to 'may require' is acceptable but slightly reduces the authors' force.
Crowd-counting models deployed on edge devices face two shifts: inputs are degraded along the capture and transmission pipeline, while models are quantized to satisfy resource constraints.
date — pass: The originalPublished value of 2026-10-26 matches the v1 submission date. The later feed date was correctly not used.
[Submitted on 26 Oct 2026]
support — pass: All claims and excerpts match the abstract. The deployment-selection implication is grounded in the authors' own conclusion. The limitations and the generalization question are appropriately scoped and add no unsupported benefits or costs.
showing why model selection must consider corruption robustness and measured deployment cost together
media — pass: The selected media is the verified direct arXiv abstract page for this paper. This is the most direct access to the 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 referenceStoring a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
Definition referenceThe delay between sending a request and receiving the result.
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 referenceStoring a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
Definition referenceThe delay between sending a request and receiving the result.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv