{"id":"c1f2cad8-c476-4d5c-995d-1a8bf2549073","arxiv_id":"2512.24561","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A new 21,535-pair RGB-thermal visual grounding benchmark and an RGBT-VGNet baseline that beats prior models, especially in low light and on small objects.","lead":"This paper introduces RGBT-Ground, a benchmark with 21,535 paired RGB and thermal images and 38,760 objects, each with referring expressions and environment labels, for testing visual grounding in difficult real-world conditions. It also adapts 11 visual grounding models to use RGB, thermal, or both, and proposes RGBT-VGNet, which performs best, especially at night and on small objects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark reliability rests on unvalidated LVLM-generated referring expressions; no agreement or human-check rate is reported.","rationale":"The paper has genuine independent support: it reuses established aligned RGB-TIR detection datasets with existing boxes, provides a unified training/evaluation framework, and benchmarks many models. The central claim, however, is a benchmark contribution, and the new annotations are auto-generated. The reader's weakest assumption correctly identified this. I find no internal contradiction that would warrant rejection; the missing validation evidence is exactly what should gate acceptance. Therefore I keep the CONDITIONAL/UNCHANGED verdict and propose a concrete validation study.","tokens_in":43321,"tokens_out":6013,"duration_ms":64412,"concrete_test":"Take a stratified random sample of 300 instances (100 from each sub-dataset), balanced across lighting/weather/size/occlusion. Have at least three independent annotators answer for each instance: (1) Is the referring expression accurate and does it uniquely identify the target box among all objects in the image? (2) Do the scene/lighting/weather/occlusion labels match the image? Report per-item agreement and Cohen's kappa. If the unique-reference validity rate is below ~95% or pairwise kappa is below ~0.6, the benchmark's reliability claim is not supported; if it is high, the main concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"RGBT-Ground's central value is as a benchmark, so the correctness of its annotations is load-bearing. Section 3.2 and Appendix D show that all referring expressions and all scene/lighting/weather/occlusion labels are produced by the Qwen-VL API; the paper only promises a 'hierarchical random sampling strategy' for human verification, without reporting the fraction checked, the sampling rule, correction statistics, or inter-annotator agreement. Table 3 validates only that the label distribution looks physically plausible, which does not establish per-instance accuracy. If even a modest fraction of expressions are generic, hallucinated, or fail to identify the target uniquely among similar objects, then Acc@0.5 on testA/B/C and the claimed RGB-vs-TIR-vs-fusion ordering are not measuring visual grounding. Because the novelty claim is essentially 'this benchmark enables reliable RGB-T grounding evaluation,' the missing annotation-quality evidence is the weakest load-bearing assumption. A secondary consistency issue: Table 2 lists testA+testB+testC instance counts that exceed the total test count (e.g., RefFLIR 837+640+986=2,463 vs 2,104), so the 'partitioned' test subsets overlap or the counts are misreported; this should be clarified when the dataset is released.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces RGBT-Ground, a large-scale RGB–Thermal visual grounding benchmark assembled from three existing RGB-T detection datasets (FLIR, M3FD, MFAD), with 21,535 aligned RGB-TIR pairs and 38,760 instances. Each instance is annotated with a referring expression generated by the Qwen-VL LVLM, a bounding box, and multi-level labels for scene type, weather, illumination, object size, and occlusion. The authors also present a unified evaluation framework (RGBT-VG) that adapts several existing RGB-only grounding models to RGB-only, TIR-only, and RGB+TIR inputs, along with a new baseline (RGBT-VGNet) using asymmetric LoRA adaptation and language-aware cross-modal fusion. Extensive experiments across three sub-datasets, multiple splits, and over 60 configurations show that RGB+TIR fusion generally improves accuracy, that low-light and small-object subsets are hardest, and that the proposed baseline achieves the highest Acc@0.5 across all three test subsets.","tokens_in":43619,"tokens_out":3233,"duration_ms":35384,"significance":"If the benchmark is released with reliable annotations and a usable evaluation protocol, it would fill a genuine gap: current visual grounding benchmarks are almost entirely RGB-only and rarely include adverse illumination, weather, or small/occluded objects. The paper provides a plausible resource with multi-level annotations and a broad evaluation of 11 models, which is valuable for future robustness research. The proposed RGBT-VGNet baseline is simple and well-motivated, and the ablation study supports its two main components. The main uncertainty is not the experimental protocol but the quality of the LVLM-generated annotations, since the benchmark's usefulness depends on the referring expressions and environmental labels being accurate and discriminative.","major_comments":[{"comment":"The annotation pipeline generates all referring expressions and all scene/weather/illumination/occlusion labels with the Qwen-VL API, but the described 'hierarchical random sampling strategy' is never quantified. The paper reports no fraction of human-checked instances, no sampling rule, no correction statistics, and no inter-annotator agreement. Table 3 only shows that the resulting distribution of lighting/weather labels is physically plausible, which does not establish per-instance quality. Since the central claim is that RGBT-Ground is a reliable benchmark, this missing evidence is load-bearing. Please report the number/percentage of instances verified by humans, the agreement between annotators and the LVLM, and the types and rates of corrections made.","section":"§3.2, Appendix D, Table 3"},{"comment":"The reported test-subset counts are internally inconsistent. For RefFLIR, testA+testB+testC = 837+640+986 = 2,463, which exceeds the listed test total of 2,104; for RefM3FD, 1,232+1,094+1,848 = 4,174 > 3,776; for RefMFAD, 789+2,452+2,550 = 5,791 > 4,244. Either the testA/B/C subsets overlap (which would need to be stated explicitly and handled in evaluation), or the counts are erroneous. This must be clarified because every reported Acc@0.5 number for testA/B/C depends on these definitions.","section":"Table 2"},{"comment":"The abstract states that 'LoRA-based models are more robust in complex scenes', but the comparison does not isolate the effect of LoRA. The two models that use LoRA (HiVG and OneRef, both from the same group) differ from the others in backbone, architecture, and training details, so any robustness difference could be due to these factors rather than the LoRA mechanism. To support this claim, the paper would need a controlled comparison that varies only the adaptation method (e.g., full fine-tuning vs. LoRA on the same architecture). Please either provide such an ablation or soften the claim to 'models with hierarchical fine-grained modulation are more robust' or similar.","section":"Abstract, §5.2, Table 4"},{"comment":"The paper repeatedly states that 'resources, annotations, code, checkpoints, and evaluation scripts have been publicly released' but the manuscript contains no URL, repository name, hash, or any concrete pointer. For a benchmark paper, data and code availability is central to reproducibility and to the community's ability to use the benchmark. Please include a stable link and, ideally, a checksum/hash for the annotations, or a clear statement about the release timeline and access procedure.","section":"Abstract, §1, §6"}],"minor_comments":[{"comment":"The learning rate is reported as '10e-4', which is nonstandard and ambiguous (10×10⁻⁴ = 1e-3). Please use '1e-3' or '1e-4' consistently.","section":"§5.1"},{"comment":"Several typos and inconsistent notations appear: 'Acc@50' is sometimes used instead of 'Acc@0.5'; 'RGB-Thermal' vs 'RGB-TIR' vs 'RGBT' are used inconsistently; appendix figures contain 'ZTY APPENDIX' watermarks that should be removed; and the appendix text contains typos such as 'omparatively', 'dark-colered', 'Penson', and 'aprked'.","section":"Throughout"},{"comment":"The related work section discusses RGB-D grounding but does not mention any prior RGB-T or infrared visual grounding benchmark. If any such work exists, it should be cited and the 'first' claim in §1 should be qualified accordingly. If none exists, a brief sentence noting this absence would strengthen the novelty claim.","section":"§2.2"},{"comment":"In several table rows, values are formatted inconsistently (e.g., '63.7291.52' and '91.5260.97' in the RefMFAD column for HiVG-L). Please ensure proper LaTeX/CSV formatting in the final version, as these are currently unreadable.","section":"Table 4/5"},{"comment":"The appendix claims 'consistent ranking across datasets' but reports no variance or statistical analysis across the multiple runs. If only one run was performed, please state this explicitly and discuss the potential for small-sample variability on the tiny scene/weather subsets (e.g., Table 12 shows many cells with only a handful of instances).","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid benchmark contribution in terms of scope and evaluation, but the lack of quantitative annotation-quality evidence is a serious gap for a dataset paper. The counting inconsistencies and missing data/code link are also likely to be raised by other reviewers. I would be happy to reconsider a revised version that addresses these points, especially if the human-verification statistics and a public release URL are added."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely useful benchmark paper — the first RGB-Thermal visual grounding benchmark, built from existing detection data, with multi-level annotations and a unified evaluation framework. The core idea is sound and the experiments are extensive. But the annotation quality evidence is thin, and there's a counting inconsistency that needs fixing before I'd trust the numbers.\n\nWhat's new: nobody else has done 2D visual grounding with paired RGB-TIR images. The benchmark fills a real gap, and the size (38K instances, 21.5K pairs) is reasonable. The multi-level labels (scene, lighting, weather, size, occlusion) are exactly what robustness evaluation needs. The RGBT-VG framework that adapts existing models to take RGB, TIR, or both is a nice contribution — it makes comparisons fair. The RGBT-VGNet baseline with asymmetric LoRA and language-aware fusion is sensible and performs well, especially on the low-light and small-object subsets.\n\nWhere it's soft: The biggest issue is the annotations. Everything is generated by Qwen-VL and only randomly sampled for human verification. No fraction checked, no inter-annotator agreement, no correction statistics. For a benchmark, that's load-bearing. Table 3 shows the label distribution looks physically plausible, but that doesn't tell you whether a given referring expression actually picks out the target. If a modest fraction of expressions are generic or hallucinated, the reported accuracy numbers and modality comparisons could be off. The authors need to quantify annotation quality — even a 5% human-checked subset with agreement numbers would help.\n\nSecond, the test split counts don't add up. In Table 2, RefFLIR testA+testB+testC = 837+640+986 = 2,463, but the test column says 2,104. Same for the other two sub-datasets. So either the subsets overlap or the counts are wrong. That should be clarified before release.\n\nThird, minor: the abstract mentions 'Tri-Prior Fusion' but the body doesn't. That looks like an editing error. Also, the abstract says resources 'have been released' while the text says 'will be released.' No link or hash yet.\n\nNone of this is fatal — the benchmark concept is good and the experiments are broad. But for a benchmark paper, the annotation quality evidence is central and currently missing. I'd want to see the release and the annotation validation before fully trusting the results.","headline":"First RGB-Thermal grounding benchmark is a real contribution, but annotation validation and split-count inconsistencies need fixing before I'd trust the numbers.","tokens_in":44100,"tokens_out":2740,"would_cite":true,"duration_ms":25702,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper introduces RGBT-Ground, the first large-scale paired RGB–thermal visual grounding benchmark, and shows that adding thermal imagery consistently recovers accuracy lost in low light, fog, small-object, and heavy-occlusion conditions","keywords":["visual grounding","RGB-Thermal","benchmark","multi-modal","referring expressions","low-light robustness","modality adaptation","CLIP"],"falsifier":"Take a random sample of 1,000 test instances from RGBT-Ground, have multiple human annotators independently write referring expressions and assign lighting, weather, and occlusion labels, then re-evaluate the trained RGB-only and RGB+TIR models on that human-annotated subset. If the human-based difficulty ordering differs materially from the machine-label-based ordering, or if the RGB-only to RGB+TIR gap shrinks by more than a third, the benchmark's conclusions do not hold.","tokens_in":43221,"feed_emoji":"🌡️","tokens_out":7725,"duration_ms":72088,"temperature":0.7,"pith_summary":"RGBT-GroundBench claims that visual grounding, the task of localizing an object from a natural-language description, is not robustly solved by RGB-only benchmarks because they omit the low-light, adverse-weather, and small-object conditions that break appearance cues. To prove this, the paper builds the first large-scale paired RGB-Thermal grounding benchmark, with 21,535 aligned image pairs and 38,760 instances labeled for scene, lighting, weather, object size, and occlusion. Across 11 adapted models and over 60 configurations, it shows a consistent difficulty ladder—well-lit > low-light, normal-size > small, clear > occluded—and that RGB+TIR fusion beats either uni-modal input in every subset. The proposed baseline, RGBT-VGNet, tops all splits. A sympathetic reader would care because safety-critical applications like autonomous driving and embodied robotics need grounding that works when RGB fails, and this benchmark gives the community a way to measure and close that robustness gap.","feed_headline":"Thermal imagery restores visual grounding in low light and fog","feed_subtitle":"First large-scale RGB-TIR benchmark shows paired fusion beats RGB-only in every condition; new baseline tops all splits.","key_machinery":"The aligned RGB-TIR image pair is the load-bearing object: thermal imagery is illumination-invariant, so it supplies stable contours exactly where RGB noise and darkness erase appearance. The paper's baseline makes this work via Asymmetric Modality Adaptation (AMA), which gives the thermal encoder a higher-rank LoRA than the RGB encoder, and Language-Aware Visual Synergy (LAVS), which uses the referring expression as a query to weight which modality's features to trust. The evaluation protocol itself—unified data pipeline, three modality settings, identical training—is what converts the dataset into a benchmark rather than a collection.","core_discovery":"The paper's central claim is that RGB-thermal visual grounding is feasible and consistently more robust than RGB-only grounding under exactly the conditions that matter for real deployment. It constructs RGBT-Ground by re-annotating three public paired RGB-TIR detection datasets, using a large vision-language model to generate referring expressions and multi-level environmental labels, then verifying a random sample. The evaluation shows that RGB-only models degrade sharply in weak light, fog, and heavy occlusion, while RGB+TIR models recover most of the gap. The proposed baseline, RGBT-VGNet, uses asymmetric LoRA adaptation (a higher-rank adapter for the thermal branch to close the RGB-pret","pith_inferences":["The difficulty ordering may partly reflect the annotation model's tendency to generate more complex expressions in hard scenes; a human-annotation study on a subset could separate true perceptual difficulty from annotation complexity.","The benchmark's value would be strengthened by a per-condition analysis—already present in the appendix—that isolates when thermal alone beats RGB alone; that result is not the paper's headline claim but is directly extractable.","The same unified framework could be extended to other paired modalities (e.g., RGB-depth) and to downstream tasks like referring segmentation and VQA, as the appendix suggests.","A practical next step is to train a per-region reliability router that decides when to trust RGB versus TIR; the baseline does this implicitly through language-guided attention but never makes it explicit."],"forward_implications":["If correct, RGB-only visual grounding should no longer be considered sufficient for safety-critical deployment; paired thermal input becomes the default recommendation.","The released benchmark provides a controlled testbed for future work, with splits isolating lighting, weather, size, and occlusion as independent difficulty axes.","The asymmetric LoRA finding suggests a general principle for adapting pre-trained RGB vision encoders to any non-RGB modality.","The roughly 10% average improvement from RGB+TIR over the best uni-modal model across all subsets makes multi-modal grounding a clear target for future research.","The baseline's code and checkpoints enable direct comparison, so progress can be measured on the same splits."],"fun_headline_variants":["Thermal fusion boosts visual grounding in low light and fog","New RGB-TIR benchmark reveals fusion outperforms RGB-only","RGBT-GroundBench: Thermal cues rescue visual grounding","Fusing thermal data makes visual grounding robust to darkness","First large-scale RGB-TIR benchmark for visual grounding"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The referring expressions and all lighting, weather, occlusion, and size labels were generated by a large vision-language model and only hierarchically random-sampled for human verification, with no reported inter-annotator agreement or human-checked fraction; if those labels are noisy or biased, the reported difficulty orderings and the size of the thermal advantage could be artifacts.","fun_headline_variants_meta":{"raw":{"variants":["Thermal fusion boosts visual grounding in low light and fog","New RGB-TIR benchmark reveals fusion outperforms RGB-only","RGBT-GroundBench: Thermal cues rescue visual grounding","Fusing thermal data makes visual grounding robust to darkness","First large-scale RGB-TIR benchmark for visual grounding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000393,"raw_usage":{"total_tokens":1940,"prompt_tokens":821,"completion_tokens":1119,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":1040}},"tokens_in":565,"tokens_out":1119,"duration_ms":8857,"temperature":1.0,"reasoning_tokens":1040,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T13:18:11.943399+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of 1,000 test instances from RGBT-Ground, have multiple human annotators independently write referring expressions and assign lighting, weather, and occlusion labels, then re-evaluate the trained RGB-only and RGB+TIR models on that human-annotated subset. If the human-based difficulty ordering differs materially from the machine-label-based ordering, or if the RGB-only to RGB+TIR gap shrinks by more than a third, the benchmark's conclusions do not hold.","supporting_citations":[],"review_version":1}