IR275K: A Benchmark for Infrared Multi-Frame Super-Resolution Toward Efficient Remote Sensing

Changxin Wang; Heyang Wang; Hongxing Qi; Hongyi Chen; Jianyu Wang; Jie Deng; Junkai Shen; Xudong Zhang; Zhiping He

IR275K: A Benchmark for Infrared Multi-Frame Super-Resolution Toward Efficient Remote Sensing

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper introduces IR275K, a 594-sequence benchmark that stress-tests infrared multi-frame super-resolution, and reports a 1.53 dB architectural lesson: SSM-based cross-frame fusion needs explicit spatial coordinates.

desk verdict IR275K is a genuinely useful benchmark resource, but the paper's own evaluation protocol is violated by using the test set for model selection, so treat the CGMamba numbers with caution. read the letter →

arxiv 2607.22380 v2 pith:6VIEN537 submitted 2026-07-24 cs.CV

Jie Deng , Heyang Wang , Changxin Wang , Junkai Shen , Hongyi Chen , Zhiping He , Hongxing Qi , Xudong Zhang

show 1 more author

Jianyu Wang

This is my paper

classification cs.CV

keywords infraredmulti-framesuper-resolutionbenchmarkdatasetstate-spacemodelsMamba2Drotarypositionencodingcross-framefusionremotesensingefficient

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that infrared multi-frame super-resolution has lacked a common testbed, because existing benchmarks are built for visible light or single-frame settings and ignore thermal contrast, sensor noise, and platform motion. IR275K supplies one: 594 infrared video sequences, split at sequence level, with an intentional gap between training and test conditions in contrast–noise coupling, temporal noise level, and inter-frame displacement, so a good score requires robustness to the failure modes real infrared payloads face. The paper further claims that a lightweight state-space model, CGMamba, which fuses neighboring frames by gating them with the central frame's features, works only when each token carries explicit 2D coordinates: removing 2D rotary position encoding costs 1.53 dB and produces grid artifacts. If true, this defines a design principle for efficient infrared MFSR: implicit fusion must be spatially anchored.

What carries the argument

The two central objects are the IR275K benchmark itself and the CGMamba reference model. IR275K's partition design is the machinery that enforces the benchmark's claim: sequence-level splits built around three stress tests—coupled low-contrast/high-noise, elevated test-time temporal noise, and larger inter-frame displacement—so that evaluation answers a robustness question, not just reconstruction quality. CGMamba's machinery is the combination of 2D rotary position encoding (2D RoPE), which injects absolute pixel coordinates into the Mamba scan's rotation angles, and Center-Guided CrossMamba (CGCM), which uses the central frame as a query to gate features from neighboring frames under the s

What would settle it

Build a paired test set where real infrared video is captured simultaneously by a low-resolution and a high-resolution detector, or where a measured optical blur and noise model are convolved into the native frames; then train and evaluate under the IR275K protocol. If methods re-rank substantially, or if CGMamba's 1.53 dB drop without 2D RoPE disappears, the benchmark's degradation model and the spatial-anchoring claim are falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a reproducible evaluation foundation and an architectural reference can together advance infrared multi-frame super-resolution. The benchmark IR275K contains 594 sequences (275,196 frames at 640×512) with sequence-level train/validation/test splits that deliberately shift contrast, noise, and motion distributions so the test set probes low-contrast, high-noise, large-displacement conditions that random splits would hide. The reference model CGMamba (10.90M parameters, 112.14G FLOPs) reaches 33.19 dB PSNR, outperforming two infrared single-image SR baselines by 0.35–0.52 dB at far lower cost. The load-bearing architectural finding is that removing 2D RoPE fro

Load-bearing premise

The benchmark treats the native 640×512 infrared frames as genuine high-resolution ground truth and uses bicubic downsampling at ×4 as the only degradation model, so if real sensor blur, noise, or compression is already present in those frames, the scores may not reflect operational infrared super-resolution performance.

Editorial extensions

If this is right

If IR275K is adopted, future infrared MFSR papers can be compared on accuracy and cost under identical stress conditions, removing the current private-dataset fragmentation.
A model that scores well on IR275K retains quality under noisier, lower-contrast, larger-displacement test splits than training—directly matching satellite payload drift and platform vibration scenarios.
The CGMamba result implies that SSM-based implicit fusion architectures for infrared should embed explicit spatial coordinates rather than rely only on content-based features.
The benchmark's protocol (three input frames, ×4 bicubic, single-channel PSNR/SSIM plus runtime/FLOPs) gives a fixed testbed for measuring the accuracy–efficiency tradeoff that matters for downlink-constrained payloads.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the benchmark is adopted, stress-stratified reporting (splitting test scores by contrast, noise, and motion bins) would become straightforward, letting the community see exactly which regime each method fails in—something the paper lists as future work.
The spatial-anchoring finding likely generalizes to any implicit fusion model, not just SSMs: under weak texture, cross-frame gating without coordinate priors may misassign features in visible-light burst SR too, though the paper only demonstrates it on infrared.
Replacing bicubic downsampling with measured sensor point-spread functions and noise fields in a future version would let the community quantify the synthetic-to-real gap on the same data foundation—the paper commits to this as a next step.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Worth a look, but with a caveat. The benchmark itself is a real contribution: 594 infrared sequences, fixed sequence-level splits, three stress dimensions (contrast/noise coupling, noise domain gap, motion), and a stated plan to release code and data. That fills a genuine gap — infrared MFSR has been evaluated on private data. The stress-test partition design is thoughtful; the motion shift (train median 0.36 px, test 0.56 px, with a long tail) is exactly the kind of distribution shift that random splits hide.

The CGMamba part is secondary but not bad. 10.9M params, 112 GFLOPs, solid efficiency reporting. The ablation story is interesting: CGCM without 2D RoPE collapses by 1.53 dB with grid artifacts, while RoPE alone does nothing. That mutual-dependency result is worth noting. The comparison against SISR baselines is honest — they don't claim to beat MFSR methods.

The soft spot is the one the stress-test note flags: every ablation and depth selection is done on the test set. Section V-B says so explicitly. The validation split exists but is never used. That leaks test information into the 33.19 dB number, the depth choice (3,3,3,3,3), and the RoPE-criticality claim. It's fixable — re-run model selection on validation — but as written, the benchmark's central promise of a held-out test set is not kept. This is the biggest issue and it's the authors' own protocol violation, not a quibble.

Two smaller issues. The degradation model is purely bicubic ×4 downsampling of the native frames; the authors acknowledge this in Sec. III-G, so it's a stated boundary, not a hidden one. And despite designing stress shifts, they don't report stress-stratified results — only aggregate test PSNR — so we don't yet see whether the stress partitions actually change model rankings. That weakens the demonstrated value of the stress design, though not the resource itself.

Who is this for? People working on infrared SR or efficient MFSR will want the benchmark. The architecture is a useful reference point, not a SOTA claim. It deserves peer review — a serious referee can push for the test-set fix and for MFSR baselines. I'd read the revised version if they address the leak.

Referee Report

3 major / 4 minor

Summary. The paper introduces IR275K, a benchmark of 594 infrared video sequences (275,196 frames) with sequence-level train/validation/test splits and a ×4 multi-frame super-resolution protocol. The splits are explicitly designed as stress tests: test sequences are noisier and have larger inter-frame displacement than training sequences, while thermal contrast is preserved. The paper also presents CGMamba, a lightweight SSM-based MFSR architecture combining 2D rotary position encoding (2D RoPE) with center-guided cross-Mamba (CGCM) fusion, and reports 33.19 dB PSNR on the IR275K test set, outperforming two infrared SISR references at lower computational cost. The central architectural claim is that explicit spatial anchoring (2D RoPE) is critical for stable SSM-based cross-frame gating, supported by an ablation in which removing 2D RoPE causes a reported 1.53 dB drop and grid-like artifacts.

Significance. If the benchmark is used as specified and its results are independently reproducible, IR275K would fill a real gap: a standardized infrared MFSR evaluation suite with controlled train/evaluation shifts in noise and motion, plus public code, fixed splits, and evaluation scripts. The stress-test design is well motivated and the dataset statistics are internally consistent. The CGMamba architectural probe, if valid, would provide a useful design principle for spatially grounded implicit fusion. However, the current evaluation protocol is compromised by test-set model selection, the numerical basis for the headline 1.53 dB claim is inconsistent with the ablation table, and the comparison lacks any MFSR baseline. These issues must be resolved before the benchmark and architectural claims can be accepted.

major comments (3)

[Sec. V-B, Tables II and III] All ablations and depth selection are evaluated on the test split: 'All experiments use identical training configurations and are evaluated on the test set for the ×4 task.' The validation split (Sec. III-F, Table I) is never used for model selection. Consequently the reported 33.19 dB, the chosen depth (3,3,3,3,3), and the 2D RoPE ablation are all selected using test labels, so they are not independent estimates. This violates the held-out evaluation protocol that a benchmark must enforce and makes the central reproducibility claim unsubstantiated. The fix is to perform structural ablations and depth selection on the validation split, freeze the configuration, and only then evaluate the final model once on the test split; report both validation and final test numbers.
[Sec. V-C, Table IV] The quantitative comparison includes bicubic interpolation and two infrared SISR methods only; there is no multi-frame baseline despite IR275K being an MFSR benchmark. Outperforming SISR references by 0.35–0.52 dB does not establish that CGMamba is an effective multi-frame method or that the benchmark differentiates MFSR approaches. Add at least one representative explicit-alignment MFSR baseline (e.g., DCUNet [34] or VSRM [37]) trained and evaluated under the same IR275K protocol; otherwise the 'accuracy–efficiency evaluation of infrared MFSR methods' claim is not demonstrated.
[Abstract; Sec. V-B; Conclusion] The abstract, introduction, and conclusion state that removing 2D RoPE from CGCM causes a 1.53 dB drop, but Table II shows Full Structure at 33.19 dB versus +CGCM (w/o 2D RoPE) at 30.96 dB, i.e., a 2.23 dB drop. The value 1.53 dB is the drop from Base (32.49) to +CGCM (30.96), not the cost of removing RoPE from the full model. Please correct the wording or the calculation; the architectural claim currently rests on an inconsistent number.

minor comments (4)

[Sec. III-F] The bicubic degradation model is stated, but the exact kernel, anti-aliasing filter, and padding settings are not specified. These details are needed for exact reproducibility of the LR/HR pairs.
[Sec. III-G] Report the input resolution at which FLOPs and runtime are computed. The current statement (RTX 4090, batch 1, average over test set) is ambiguous without specifying the input patch size used in Table III and Table IV.
[Sec. V-C] Because SISR baselines consume one input frame while CGMamba consumes three, clarify whether runtime and FLOPs are reported per input frame or per video sequence. This affects the efficiency comparison.
[Sec. III-B] The term 'three stress tests' is slightly misleading: the contrast–noise coupling is preserved across splits rather than shifted, while the explicit train/test shifts are in noise level and motion magnitude. Consider rephrasing to clarify which factors are shifted and which are coupled.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark is an external artifact, and CGMamba's results are empirical measurements, not derivations from fitted inputs.

full rationale

I examined the derivation chain for the benchmark claim and the CGMamba architectural claim. IR275K is constructed from three public infrared sources (AnyTSR++, Science Data Bank), filtered by curation criteria, and split at sequence level; the PSNR/SSIM numbers for CGMamba and baselines (Tables II–IV) are measured under this fixed protocol. No equation reduces to a fitted value that is later reported as a prediction: Eq. (1)–(2) define 2D RoPE phase angles, Eq. (3)–(6) define the SSD-based cross-frame gating, and Eq. (7) is the Charbonnier training loss. The central architectural finding—removing 2D RoPE causes a 1.53 dB drop—is an empirical ablation result (Table II, Fig. 8), not a parameter fitted to data and then renamed as a discovery. No load-bearing self-citation exists: references to Mamba, SSD, RoFormer, and Mamba-3 are external prior work, and the dataset/benchmark is not justified by a self-citation chain. The test-set protocol concern raised in the review—'All experiments use identical training configurations and are evaluated on the test set for the ×4 task' (Sec. V-B), meaning depth selection (Table III) and structural ablations (Table II) are decided from test-set PSNR—is a real evaluation-integrity issue (test information can leak into model/hyperparameter selection), but it is not circularity in the sense defined here: the reported numbers are still externally measured, would change under different splits or hyperparameters, and do not make the benchmark's definition equivalent to its conclusions. The paper's own limitation statements (Sec. III-G and Conclusion) also explicitly acknowledge that the benchmark uses bicubic-only degradation and has not yet incorporated real-sensor effects, which further supports that the evaluation claims are contingent empirical claims rather than tautological constructions.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on hand-chosen architecture hyperparameters, qualitative curation thresholds, and domain assumptions about synthetic degradation and ground-truth validity. No new physical entities are introduced. The most consequential free parameters are the depth configuration and the split-exclusion criteria, because they directly determine the headline numbers and the benchmark population.

free parameters (4)

CGMamba depth configuration = (3,3,3,3,3)
Selected by test-set evaluation across Table III; directly determines the reported 33.19 dB PSNR and 112.14 G FLOPs.
embedding dimension / SSM state dimension = embed_dim=128, cross_d_state=32
Chosen by hand in Sec. V-A; sets model capacity, parameter count, and FLOPs, and is not justified by validation.
learnable mixing coefficient α = learned (not reported)
Eq. (2) blends Mamba's content angle with 2D RoPE; the learned value is not analyzed, but the architecture's behavior depends on it.
sequence exclusion thresholds = not quantified
Sec. III-D removes static, severely blurred, and extreme-displacement sequences using qualitative criteria; these choices shape the entire benchmark population and splits.

assumptions (5)

domain assumption Bicubic ×4 downsampling is an adequate degradation model for infrared MFSR evaluation
Stated in Sec. III-F/III-G; the benchmark's HR/LR pairs are all generated this way, and the authors acknowledge real-sensor effects are excluded.
domain assumption Native 640×512 infrared frames are valid high-resolution ground truth
The curation uses the source frames as HR targets without accounting for sensor blur, noise, or compression already present in them (Sec. III-F).
domain assumption A test split with elevated noise and larger motion than training is the right evaluation criterion
The controlled difficulty shifts in Sec. III-B assume that harder test conditions are the primary way to expose deployment robustness.
ad hoc to paper The SSD expansion Eq. (3) supports the decoupled cross-frame Q/K/V construction in Eq. (6)
Sec. IV-C uses the SSD recurrence as a basis for center-guided cross-Mamba, but the decoupling of Q from the central frame and K,V from the neighbor frame is an architectural postulate, not a theorem proven in the paper.
domain assumption The public source datasets (AnyTSR++, Science Data Bank) are representative of operational infrared sensing
Sec. III-C relies on three public sources for diversity, but no evidence is given that their capture conditions match real satellite/aircraft payloads.

how reviews work

0 comments

Cite this review

Pith. "Pith review of IR275K: A Benchmark for Infrared Multi-Frame Super-Resolution Toward Efficient Remote Sensing." pith.science (2026). https://pith.science/paper/6VIEN537

@misc{pith2026260722380,
  author       = {Pith},
  title        = {Pith review of: IR275K: A Benchmark for Infrared Multi-Frame Super-Resolution Toward Efficient Remote Sensing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6VIEN537}},
  note         = {Machine review of arXiv:2607.22380}
}

read the original abstract

Efficient processing is becoming increasingly important in infrared remote sensing, where satellite constellations produce large volumes of observations under constrained detector resolution, power, and downlink bandwidth. Multi-frame super-resolution (MFSR) offers a software-based route to spatial enhancement, but its evaluation in infrared sensing remains fragmented across private datasets and ad-hoc protocols. Existing benchmarks do not explicitly capture the thermal contrast, sensor noise, weak texture, and platform-induced frame-to-frame variation that characterize infrared video. We introduce IR275K, a curated benchmark containing 594 infrared video sequences and 275,196 frames. It provides sequence-level train/validation/test splits and a reproducible X4 evaluation protocol. As an initial architectural probe, we further evaluate CGMamba, a lightweight state-space model with 10.90M parameters and 112.14G FLOPs. CGMamba combines 2D rotary position encoding (2D~RoPE) with center-guided cross-Mamba (CGCM) fusion for implicit multi-frame reconstruction. It achieves 33.19dB PSNR, outperforming infrared single-image super-resolution references by 0.35--0.52~dB at substantially lower computational cost. Ablation results show that removing 2D~RoPE from CGCM causes a 1.53dB drop and severe grid-like artifacts. This indicates that explicit spatial anchoring is critical for stabilizing SSM-based cross-frame gating under infrared conditions. IR275K provides a reproducible foundation for accuracy--efficiency evaluation of infrared MFSR methods, while the architectural analysis offers a concrete starting point for spatially aware SSM design under resource-constrained infrared sensing. Dataset and evaluation resources are available at: https://github.com/InfraRecon7/IR275K.