{"id":"dfcc89ce-cc4b-4191-a84f-8ead2c8bf458","arxiv_id":"2507.20976","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A multi-stage diffusion-based framework that generates labeled synthetic aerial images from weak image-level labels improves cross-domain vehicle detection AP50 over prior adaptation methods.","lead":"This paper builds a pipeline that generates synthetic aerial images and bounding-box labels, then uses them to train vehicle detectors that transfer to unseen geographic regions using only image-level labels. It reports large AP50 gains over source-only, unsupervised, and weakly supervised baselines, and introduces two new aerial datasets from New Zealand and Utah.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central assumption—that cross-attention maps transfer localization from source to target domains—is never directly measured; the method's gains could come from target-like synthetic images with noisy labels rather than from the attention-based labeling.","rationale":"I agree with the reader's weakest assumption. The headline numbers depend on the pseudo-label chain in Algorithm 1: the final detector is trained on synthetic target images labeled by a detector FA that was trained on source attention maps. This is the one place where an unmeasured assumption carries the whole method; if the synthetic target labels are substantially wrong, the experiments no longer test the proposed mechanism. I checked the 42.36 px pseudo-bounding-box derivation in Appendix D.1 and it is consistent with the stated decision-circle equivalence, so I do not treat that as a load-bearing flaw. I also considered diffusion memorization as a possible confound, but the more fundamental and directly testable unknown is label accuracy. The concern is not that the assumption is false, only that it is asserted rather than demonstrated, and the manuscript's own limitations about 8x8 attention maps and overlapping objects show plausible failure modes. Because this is addressable by a direct annotation study plus a simple baseline, the CONDITIONAL verdict remains appropriate.","tokens_in":20932,"tokens_out":7111,"duration_ms":86537,"concrete_test":"Manually annotate vehicle centers in a random sample of N=200–300 synthetic target images xGT using the paper's 12 px decision-circle criterion, then compare the pipeline's automatic labels yGT against this manual ground truth, reporting precision, recall, and AP50. Also run a fixed-pipeline ablation that trains the final detector on synthetic target images labeled by FS directly applied to xGT, skipping the attention-map transfer. If the attention-derived labels have low precision/recall, or if the direct-FS labels match or beat them, the claim that style-less cross-attention maps enable label transfer is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing step is the transfer of localization from source to target attention maps (Sec. 3.3, Algorithm 1 steps 5–7). FS, trained on real source images, produces pseudo-labels for synthetic source images; FA is trained on those labels paired with source attention maps; then FA labels target attention maps to produce the synthetic target labels, and the final detector is trained on synthetic target images. This chain is valid only if FS is accurate on synthetic source images and if FA transfers across attention-map domains. The paper does not measure either condition. Section 3.3 asserts that cross-attention maps 'contain less style information than RGB images' and therefore support transfer, but no experiment compares source and target attention maps, and no experiment reports the accuracy of the resulting target labels against manual annotations on synthetic target images. The paper's own Sec. 6 concedes that attention maps are 8x8 and that overlapping objects make labeling difficult, so failure modes are plausible. Without direct measurement, the Table 1 gains could be explained by training on target-like synthetic RGB images with low-quality labels rather than by the proposed attention-based labeling mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses cross-domain vehicle detection in aerial imagery under a weak-supervision setting: fully labeled source images (LINZ or DOTA) and image-level target labels (UGRC). The proposed AGenDA pipeline fine-tunes a latent diffusion model on source and target images with domain-specific prompts, synthesizes 10k source and 10k target car images plus 10k empty target images, and labels the synthetic target images by a three-stage transfer: a detector FS trained on real source images pseudo-labels synthetic source images; a second detector FA is trained on stacked source cross-attention maps with those pseudo-labels; FA then predicts boxes on stacked target attention maps; the final detector is trained on synthetic target RGB images with these pseudo-labels and evaluated on the real target test set. Stacked attention maps combine the category word and two learnable tokens (foreground and background), with total-variation regularization. A classifier refinement module selects high/low-confidence labels. Experiments compare four detectors against source-only, open-set, UDA, and weakly supervised baselines, reporting AP50 gains (e.g., YOLOv8: 75.4% LINZ→UGRC and 75.7% DOTA→UGRC). The paper also introduces two new aerial datasets.","tokens_in":21130,"tokens_out":10703,"duration_ms":123416,"significance":"The proposed use of diffusion models for annotated aerial synthetic data is timely and the experimental scope is unusually broad: four detector backbones, two domain shifts, and multiple baseline families. The consistent double-digit gains over strong weak-supervision baselines (H2FA R-CNN, OCUD) are impressive if reproducible. However, the paper's central mechanism—the transfer of localization knowledge through grayscale cross-attention maps—is asserted rather than verified, and the evaluation rests on single-run AP50 numbers without uncertainty quantification. The new LINZ and UGRC datasets are a useful community resource, but they are not made available, and no code is released. The value of the contribution depends on closing these gaps.","major_comments":[{"comment":"All AP50 values in Table 1 (and the ablations in Table 2 and Figure 5) are reported as single runs with no error bars, seeds, or significance tests. The reported advantage over the best weakly-supervised baseline is 6.2 and 7.2 percentage points, while several internal ablations differ by as little as 1.2 points (Table 2, DOTA→UGRC). Given stochasticity in detector and diffusion training, these differences are within the range where ranking could change across seeds. Please report mean±std over at least 3–5 seeds for the main comparisons and ablations, and state whether the headline improvements are statistically reliable.","section":"§5.2, Table 1"},{"comment":"The core assumption of the label-transfer chain—that cross-attention maps are domain-invariant and localization-accurate enough to transfer a detector from source to target—is never directly measured. Step 5 assumes FS produces reliable pseudo-labels on synthetic source images; step 7 assumes FA trained on source attention maps transfers to target attention maps; and Section 3.3 justifies this by stating that attention maps 'contain less style information than RGB images,' but no experiment compares the source/target attention distributions or evaluates the resulting target pseudo-labels. Section 6 even concedes that the 8×8 attention resolution and overlapping objects make labeling difficult. Please add (i) quantitative label-quality evaluation of yGT against manually annotated boxes on synthetic target images, and (ii) an ablation that trains the final detector on synthetic target images labeled by an alternative mechanism (e.g., FS applied directly to synthetic target RGB, or attention thresholding without the FA transfer). Without these, the Table 1 gains cannot be attributed to the proposed attention-based transfer; they could be due to target-like synthetic RGB images with noisy labels.","section":"§3.3, Algorithm 1"},{"comment":"The paper replaces the stated decision-circle criterion (12 px radius) with a 42.36 px pseudo-bounding box and claims this is 'functionally equivalent' with 'minimal error,' but the error is never quantified. The isocontour IoU=0.5 region only matches the quarter-circle area, not its shape, so some centers inside the decision circle can be counted as false positives and vice versa. Because every number in Tables 1–3 and Figure 5 depends on this substitution, please report the worst-case or expected discrepancy between AP50 computed with the pseudo-box and AP computed directly with the decision circle, or include both metrics.","section":"§5.1 and §D.1"},{"comment":"The paper introduces two new datasets and a multi-stage pipeline but does not provide code, model weights, or a data release link (the project page URL is given, but no artifacts are described). Given the large number of implementation choices (Stage 1/2/3 hyperparameters in Table 4, thresholds, synthetic image counts), the central experimental claims are not independently reproducible without these artifacts. Please release code and the datasets (or a documented access procedure) at the project page.","section":"Datasets and reproducibility"}],"minor_comments":[{"comment":"The row for YOLOv8 cites reference [15] (YOLOv5); it should cite [44]. Please check all table citations for consistency.","section":"Table 1"},{"comment":"The phrase 'use the well-trained detector FA to test the target domain cross-attention maps' should be 'to test on the target domain cross-attention maps'.","section":"§3.3"},{"comment":"The use of the phrase 'functionally equivalent' for the pseudo-bounding-box approximation is too strong without a quantified error bound; the wording should be softened or supported by the error analysis requested above.","section":"§5.1"},{"comment":"The description says all training stages scale image resolution to 128×128 because YOLOv5 requires multiples of 32. Since the datasets are introduced at 112×112, please clarify whether evaluation is also performed at 128×128 and whether the 112×112 sampling size is preserved in any stage.","section":"Appendix D.2"}],"recommendation":"major_revision","confidential_remarks":"This is a strong empirical CVPR-style paper, but for a journal I would want the transfer mechanism validated and the reproducibility artifacts in place before acceptance. I am not recommending rejection: the method is plausible, the experiments are extensive, and the requested additions (seeded runs, label-quality analysis, and code/data release) are feasible within a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nYou should know two things about this paper. First, it reports a genuinely new pipeline for cross-domain vehicle detection in aerial imagery: fine-tune a latent diffusion model on source and target domains, generate synthetic images, extract stacked cross-attention maps, and use a detector trained on those maps in the source domain to label target attention maps, then train a final detector on synthetic target images with those pseudo-labels. The experiments are extensive, covering four detectors, and the authors contribute two new annotated aerial datasets from New Zealand (LINZ) and Utah (UGRC). Second, the load-bearing assumption—that cross-attention maps are domain-invariant enough to transfer localization from source to target—is asserted but never measured. If it fails, the reported gains could simply come from training on target-like synthetic images with noisy labels.\n\nWhat the paper does well: the two-stage diffusion fine-tuning with learnable foreground and background tokens is a sensible extension of prior work, and the ablation showing stacked maps outperform the single 'car' map is informative. The two datasets, with point annotations and spatially disjoint train/test splits, are a real contribution. The comparison against open-set detectors and VLLMs is honest, even if those baselines are predictably weak.\n\nThe soft spots are real but addressable. There is no code, no data link, and no error bars or multiple seeds, so the large AP50 differences cannot be checked for significance. The pseudo-bounding box metric converts a 12-px decision circle into a 42.36-px square by matching the area of the IoU≥0.5 region; that guarantees equal area but not equal per-point decisions, so 'functionally equivalent' is overclaimed. Most importantly, the cross-domain transfer step is never validated: no experiment measures label accuracy on synthetic target images, and no comparison of attention-map statistics across domains is reported. Given the authors' own limitation about 8x8 attention maps and overlapping objects, this gap matters.\n\nThe stress-test concern lands. The fix is straightforward: manually annotate a subset of synthetic target images, report precision/recall of the auto labels, and compare attention-map distributions across source and target. Also add error bars or repeated runs.\n\nThis paper deserves a serious referee. It is not ready as-is, but the core idea and the datasets are worth developing. I would recommend conditional acceptance with major revisions, contingent on the missing measurements and on releasing code and data.","headline":"Novel diffusion-based cross-domain labeling pipeline with strong reported gains, but the key transfer assumption is never directly measured and no artifacts are released.","tokens_in":21710,"tokens_out":5340,"would_cite":true,"duration_ms":56727,"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":"A diffusion-based pipeline generates synthetic aerial images and derives their labels from stacked cross-attention maps, lifting cross-domain vehicle detection to 75.7 AP50 without target bounding-box annotations.","keywords":["cross-domain object detection","aerial vehicle detection","latent diffusion models","cross-attention maps","weakly supervised domain adaptation","synthetic data augmentation","small object detection","domain shift"],"falsifier":"Train a binary classifier to distinguish stacked source attention maps from stacked target attention maps; if the classifier performs well above chance, the attention maps carry domain information and the label-transfer chain is suspect. A direct check would compare the pipeline's pseudo-labels on synthetic target images against human annotations of the same images and require label accuracy at least as high as the final detector's AP50.","tokens_in":20683,"feed_emoji":"🚗","tokens_out":8577,"duration_ms":95214,"temperature":0.7,"pith_summary":"The paper claims that you can adapt a vehicle detector to a new aerial domain without bounding-box annotations by generating synthetic images of the target domain and labeling them automatically. The labels come from cross-attention maps inside a fine-tuned text-to-image diffusion model: three stacked maps for a car concept, a learned foreground token, and a learned background token are turned into detector inputs, and a detector trained on synthetic source attention maps transfers to synthetic target attention maps. With only image-level labels for the target, the pipeline then trains the final detector on synthetic target images and their pseudo-labels. Across two domain shifts, this method reports AP50 gains over source-only training of 4–23%, and best scores of 75.4% and 75.7% with YOLOv8, beating weakly supervised and unsupervised adaptation baselines. The paper also contributes two new annotated aerial datasets, LINZ and UGRC.","feed_headline":"Diffusion-generated aerial images lift vehicle detection to 75.7 AP50","feed_subtitle":"Training on synthetic aerial images and attention-derived labels boosts Utah detection without target boxes.","key_machinery":"The load-bearing object is the stacked (multi-channel) cross-attention map of a fine-tuned latent diffusion model. For each synthetic image, the paper extracts cross-attention maps for the word 'car', for a learnable foreground token [V1], and for a learnable background token [V2]/[V3], normalizes them, and stacks them into a three-channel 'style-less' representation. These maps are what lets labels travel from source to target: a detector trained on stacked maps of labeled synthetic source images is applied to stacked maps of unlabeled synthetic target images, producing boxes that are then attached to the corresponding synthetic target RGB images. Two fine-tuning stages align the diffusion model with aerial views and with the learned tokens, using a total-variation loss that pushes the foreground token map toward the category map and the background token maps toward its complement.","core_discovery":"The paper's central claim is that a fine-tuned latent diffusion model can serve as a data engine for cross-domain aerial vehicle detection: it can synthesize target-domain aerial images together with usable bounding-box labels, even though the model never sees target-box annotations. The label-generation mechanism is a chain of transfers through cross-attention maps. A detector is first trained on real source images with full boxes; it labels synthetic source images; a second detector is trained on the stacked cross-attention maps of those synthetic source images; and that second detector labels the stacked cross-attention maps of synthetic target images. The final, deployable detector is trained on synthetic target RGB images with those pseudo-labels and evaluated on real target images. The reported results support the claim that this label-transfer chain closes much of the domain gap.","pith_inferences":["If the cross-attention maps are not as domain-invariant as assumed, the reported gains would mostly reflect having more training data and better synthetic images rather than successful label transfer; measuring source-vs-target attention-map shift would separate those explanations.","The pipeline should carry over to other small-object categories and other overhead imaging modalities, but the 112×112 pixel crop and 8×8 attention resolution set a floor on object size, so gains would shrink for extremely small or densely overlapping objects.","Because the evaluation metric is center-localization within a 12-pixel decision circle, the reported AP50 numbers may not reflect tight-box quality; users who care about precise footprint should request additional IoU thresholds."],"forward_implications":["On the two tested shifts, LINZ→UGRC and DOTA→UGRC, adding synthetic target images with attention-derived labels raises AP50 over source-only training by 4–23 percentage points across Faster R-CNN, YOLOv5, YOLOv8, and ViTDet.","The pipeline beats prior weakly supervised cross-domain detectors by 6–10 points and unsupervised domain adaptation by 7–40 points on the same shifts, using only image-level target labels.","Open-set detectors and large vision-language models perform poorly on aerial vehicles in their zero-shot settings, which positions synthetic data generation as a more practical adaptation route when target boxes are unavailable.","The label-refinement classifier makes the labeling robust to confidence-threshold differences across datasets, removing a threshold-tuning step that would otherwise be needed per target domain."],"supporting_citations":[{"why":"Supplies the latent diffusion backbone that is fine-tuned on source and target aerial images and used for synthesis.","marker":"[46]"},{"why":"Establishes that cross-attention maps in text-to-image diffusion localize the mentioned concept, the basis for using them as labels.","marker":"[54]"},{"why":"Introduces learnable tokens placed before target concepts to sharpen cross-attention maps, which the paper adapts for foreground and background tokens.","marker":"[37]"},{"why":"Shows how unique identifiers in prompts can be fine-tuned to represent specific concepts and domains, enabling the source-target domain prompts.","marker":"[47]"},{"why":"Demonstrates deriving pixel-level labels from diffusion cross-attention maps, the precedent this pipeline extends to cross-domain detection.","marker":"[60]"},{"why":"Uses attention maps from diffusion models to label generated object-detection data, a related mechanism the method builds on.","marker":"[70]"}],"fun_headline_variants":["Diffusion model adapts aerial vehicle detectors to unseen domains","Synthetic aerial images boost detector domain adaption without target boxes","Latent diffusion transfers labels across aerial domains for detectors","Cross-attention label transfer adapts vehicle detectors to new regions","Generative data engine closes domain gap in aerial vehicle detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that stacked cross-attention maps are nearly free of domain style, so a detector trained on synthetic source attention maps can correctly label synthetic target attention maps; the paper states this premise but runs no experiment that measures the transfer accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model adapts aerial vehicle detectors to unseen domains","Synthetic aerial images boost detector domain adaption without target boxes","Latent diffusion transfers labels across aerial domains for detectors","Cross-attention label transfer adapts vehicle detectors to new regions","Generative data engine closes domain gap in aerial vehicle detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000426,"raw_usage":{"total_tokens":2183,"prompt_tokens":949,"completion_tokens":1234,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":1152}},"tokens_in":565,"tokens_out":1234,"duration_ms":10731,"temperature":1.0,"reasoning_tokens":1152,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:04:26.345602+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a binary classifier to distinguish stacked source attention maps from stacked target attention maps; if the classifier performs well above chance, the attention maps carry domain information and the label-transfer chain is suspect. A direct check would compare the pipeline's pseudo-labels on synthetic target images against human annotations of the same images and require label accuracy at least as high as the final detector's AP50.","supporting_citations":[{"cited_title":"High-Resolution Image Synthesis With Latent Diffusion Models","cited_arxiv_id":null,"evidence_quote":"Supplies the latent diffusion backbone that is fine-tuned on source and target aerial images and used for synthesis."},{"cited_title":"What the DAAM: Interpreting Stable Dif- fusion Using Cross Attention","cited_arxiv_id":null,"evidence_quote":"Establishes that cross-attention maps in text-to-image diffusion localize the mentioned concept, the basis for using them as labels."},{"cited_title":"AttnDreamBooth: To- wards Text-Aligned Personalized Text-to-Image Generation","cited_arxiv_id":null,"evidence_quote":"Introduces learnable tokens placed before target concepts to sharpen cross-attention maps, which the paper adapts for foreground and background tokens."},{"cited_title":"DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation","cited_arxiv_id":null,"evidence_quote":"Shows how unique identifiers in prompts can be fine-tuned to represent specific concepts and domains, enabling the source-target domain prompts."},{"cited_title":"DatasetDM: Synthesizing Data with Perception An- notations Using Diffusion Models","cited_arxiv_id":null,"evidence_quote":"Demonstrates deriving pixel-level labels from diffusion cross-attention maps, the precedent this pipeline extends to cross-domain detection."},{"cited_title":"DiffusionEngine: Diffusion Model is Scalable Data Engine for Object Detection","cited_arxiv_id":"2309.03893","evidence_quote":"Uses attention maps from diffusion models to label generated object-detection data, a related mechanism the method builds on."}],"review_version":1}