{"id":"68932d7b-daa5-4aca-b87b-f9ae25cd3cae","arxiv_id":"2412.11582","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A new benchmark for oriented tiny object detection with the smallest object sizes to date, plus a dynamic coarse-to-fine training scheme that improves accuracy across eight datasets.","lead":"This paper introduces AI-TOD-R, a new dataset of 752,460 extremely small rotated objects in aerial images, and a training scheme called DCFL that improves detection of these tiny objects. The work gives the computer vision community a benchmark for oriented tiny object detection and a method that can be added to existing detectors without extra cost at deployment.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's 'costless improvement' claim is internally inconsistent: the Prior Capturing Block (PCB) is used at inference yet Table IV reports FLOPs identical to the baseline, so the efficiency claim needs verification.","rationale":"The reader's weakest assumption concerned the scale-invariance/discriminability of the GJSD similarity for sub-16px objects. That is a legitimate concern, but it is not the most load-bearing for the paper's central claim. The ablation in Table XIb shows GJSD only modestly outperforms KLD (59.15 vs. 57.82) and GWD (58.55), so even if GJSD degraded for extremely tiny boxes, the coarse-to-fine pipeline would still deliver most of its gains. The stronger vulnerability is the paper's explicit 'costless improvement' claim. The method text and Figure 6 describe a learnable Prior Capturing Block used during inference, yet Table IV reports exactly the same FLOPs as the baseline and Table II reports the same or lower parameter counts. This is internally inconsistent unless the PCB's cost is excluded from the profile or the baseline comparison uses a different prior density. Since 'high efficiency' is one of the three highlighted characteristics of DCFL (along with accuracy and versatility), a false or unverified costless claim would materially weaken the paper's central message. The dataset and benchmark contributions remain valuable, and the accuracy gains are plausibly real, so the appropriate outcome is still conditional acceptance with a requirement to reproduce and correct the efficiency numbers. I therefore keep the reader's CONDITIONAL verdict rather than moving to REJECT, while flagging that the efficiency claim needs concrete verification.","tokens_in":30084,"tokens_out":13127,"duration_ms":126640,"concrete_test":"Use the released code to build RetinaNet-O with and without the PCB under the SODA-A 1200x1200 setting of Table IV, and measure parameter count and FLOPs with a standard profiler (e.g., fvcore or thop). Then compare against the reported 221.90G FLOPs. If the DCFL model's FLOPs or parameter count exceeds the baseline, or if matching the reported number requires excluding the PCB from the profile, the 'costless inference' claim is refuted and Table IV must be corrected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper claims DCFL improves detection 'without adding any parameter or computational overhead during inference' (Section I, 'Costless improvement') and lists high efficiency as a central characteristic. However, the method's dynamic prior (Prior Capturing Block, Eqs. 2 and 8; Figure 6) is a learnable dilated-convolution plus deformable-conv offset branch that is part of the inference-time forward pass. Adding such a module must increase both parameter count and FLOPs. Table IV reports DCFL (RetinaNet-O) as 36.1M params and 221.90G FLOPs, while the RetinaNet-O baseline is 36.3M params and 221.90G FLOPs; Table II also shows the same or lower parameter counts for DCFL variants. This is only possible if the PCB parameters replace other parameters (e.g., by reducing the number of anchor priors) and if the FLOPs measurement omits the PCB or the added cost is so small that it vanishes at two-decimal precision. If the PCB is omitted from the FLOPs profile, the reported 'costless' claim is misleading: the method adds inference-time computation, and the claimed high efficiency is overstated. The accuracy gains in Tables II, IV, and V are not at issue; the load-bearing problem is that a core explicit claim of the paper is contradicted by its own architectural description.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper targets oriented tiny object detection and makes three contributions: a new dataset AI-TOD-R (28,036 images, 752,460 oriented instances, mean object size 10.6 px), a benchmark covering fully-supervised and label-efficient detection paradigms, and a training scheme called Dynamic Coarse-to-Fine Learning (DCFL). DCFL consists of a Prior Capturing Block (PCB) that updates anchor/prior positions and a two-stage sample assignment procedure (CPS, MPS, DGMM) intended to balance the quantity and quality of positive samples across object scales and orientations. Experiments are reported on eight datasets, including AI-TOD-R, SODA-A, DOTA-v1/v1.5/v2, DIOR-R, VisDrone, and MS COCO, with claims of state-of-the-art accuracy, high efficiency, and versatility. The dataset, benchmark, and code are promised to be publicly available.","tokens_in":1795,"tokens_out":2073,"duration_ms":107625,"significance":"If the main claims hold, AI-TOD-R is a valuable resource: it is smaller in mean object size than existing oriented object detection datasets, and the benchmark across fully-supervised, semi-supervised, sparsely-annotated, and weakly-supervised settings is informative for the community. The DCFL accuracy improvements are substantial and are tested against standard MMRotate baselines on multiple datasets; the coarse-to-fine assignment idea is a reasonable training-side mechanism for tiny objects. The paper also lists honest limitations (closed-set setting, aerial optical scenario) in the discussion. However, the central 'costless improvement' claim is not yet demonstrated, and the relationship between the baseline and DCFL prior configurations is not fully specified; these points affect how the accuracy numbers should be interpreted. The release of code and dataset is a concrete strength that supports reproducibility.","major_comments":[{"comment":"The headline 'costless improvement' claim is not substantiated by the paper's own architectural description. Section I states that DCFL improves detection 'without adding any parameter or computational overhead during inference,' and Section VI-B repeats that the accuracy improvement incurs no extra parameter or computational cost. However, the PCB is described as a learnable block that is part of the detection pipeline (Eq. 2), with an offset-prediction branch and a guided deformable convolution (Section V-B, Eq. 8, Figure 6). If this block runs at inference time, it necessarily adds FLOPs and parameters. The reported numbers in Tables II and IV are compatible with net parameter counts being unchanged or slightly lower, but no breakdown is given that isolates the PCB's contribution or shows how the FLOPs profile was computed. Please provide a direct measurement of parameters and FLOPs with and without the PCB, state explicitly whether the dynamic prior is applied at test time, and, if it is applied only during training, explain how training-time prior updates transfer to a detector that uses static priors at inference. The accuracy results may be valid, but the efficiency claim as written is currently unverified.","section":"Section I, V-B, VI-B; Equations (2), (8); Tables II, IV"},{"comment":"The baseline comparisons may not be matched in prior configuration. Section VI-C says 'we tile one prior for each feature point in all experiments,' but it is not clear whether this statement applies to all main experiments or only to the ablations on DOTA-v2. Table II reports RetinaNet-O as 36.3M parameters and DCFL(RetinaNet-O) as 36.1M, and Table IV reports 36.2M versus 36.1M; these differences suggest that the number of priors or the head structure differs between the baseline and DCFL. If the baseline uses the default MMRotate multi-anchor configuration while DCFL uses a single prior plus the PCB, then the reported gains in Tables II, IV, and V conflate the DCFL learning scheme with a change in prior density. Please specify, for each main experiment, the exact prior/anchor configuration used by both the baseline and the DCFL variant, and provide a matched comparison in which the baseline uses the same prior configuration as DCFL. The DOTA-v2 ablation in Table XIIc is a step in this direction, but it is not sufficient to establish that the main-table comparisons are apples-to-apples.","section":"Section VI-C; Tables II, IV, V"}],"minor_comments":[{"comment":"The mathematical notation in Equations (11) and (12) is garbled: the displayed expressions 'Σα = (ΣpΣg)Σ α' and 'µα = (µpµg)µ α' are typographically corrupted and should be written as the standard harmonic mean formulas for the Gaussians in the generalized Jensen-Shannon divergence.","section":"Section V-C, Equations (11)-(12)"},{"comment":"All main results are reported as single runs without standard deviations or multiple seeds. Given that the paper claims improvements of 2-8 AP points, reporting at least the mean and standard deviation over three seeds for the core comparisons (especially AI-TOD-R and SODA-A) would substantially increase confidence in the results.","section":"Tables II, IV, V; Section VI-C"},{"comment":"The ablations are performed only on DOTA-v2 val, not on AI-TOD-R. Since the paper's central motivation is oriented tiny objects, adding at least one ablation on AI-TOD-R (e.g., the individual effectiveness of CPS/MPS/DGMM and the PCB components) would directly support the causal story for the main dataset.","section":"Section VI-C"},{"comment":"The table lists 'SODA-D' but the text never defines or discusses SODA-D; please add a citation and one sentence of context.","section":"Table I"},{"comment":"The label-efficient benchmark includes only one SAOD method (Co-mining). The conclusion that SSOD outperforms SAOD should be qualified as based on a single SAOD baseline, since other SAOD methods may behave differently.","section":"Section IV-C"},{"comment":"The caption contains a run-on sentence beginning 'By separating the model's predictions...' that is not grammatical; please rephrase.","section":"Figure 7 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper's dataset and benchmark contributions are solid and the accuracy gains are plausible, but the 'costless improvement' claim is stated as a headline characteristic and is not backed by the required profiling. The possible mismatch in prior configuration between baselines and DCFL is also load-bearing for interpreting the main tables. The authors should be asked to provide direct parameter/FLOPs measurements and to clarify the prior configuration. The paper's reliance on the authors' own prior dataset AI-TOD-v2 and their CVPR 2023 DCFL method is natural for a journal extension, but the incremental novelty beyond the conference version should be made explicit in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: if you work on oriented or tiny-object detection, the AI-TOD-R dataset and benchmark are the real contribution here, not the method. DCFL goes back to the authors' CVPR 2023 paper (ref [16]), and this journal version mostly re-runs it on more datasets and plugs it into more detectors. That is fine, but referees should know the novelty is in the resource, not the algorithm.\n\nWhat the paper does well: AI-TOD-R is the smallest-scale oriented detection dataset, with a mean object size of 10.6px, and the semi-automatic annotation pipeline is sensible and transparent. The benchmark is broader than most, covering fully supervised, semi-supervised, sparsely annotated, and weakly supervised methods. The finding that label-efficient methods are competitive on tiny oriented objects is worth having. The experimental sweep is wide: eight datasets, one- and two-stage detectors, horizontal and oriented boxes. Reported gains are large and broadly consistent. The analysis of learning bias, with quantity and quality of assigned samples, is more convincing than the method equations.\n\nSoft spots, in proportion. All main results are single runs; no error bars or seeds. For a dataset benchmark paper that is tolerable, but the authors should report variance. The \"costless improvement\" language is imprecise. I checked the stress-test concern: the Prior Capturing Block is a learnable module in the inference graph, so it is not literally free. But the reported parameters for DCFL variants are equal to or lower than the baseline, which suggests the DCFL variant uses one prior per location while the baseline uses more anchors; that architecture change can offset the added compute. So the stress-test assertion that adding a module must increase FLOPs does not fully land. Still, the paper should state the prior-count change explicitly and provide a FLOPs breakdown. Ablations are all on DOTA-v2, and hyperparameters K, Q, g are tuned there; the robustness claims are supported only by small fluctuations around one operating point. Finally, the GJSD similarity is imported from larger-object settings. Scale invariance at 10px is plausible because it is a ratio-based divergence, but the paper does not re-derive it for this regime. That is the most fragile assumption, and the consistent empirical gains buy it some credibility.\n\nBottom line: this is a resource paper with a serviceable method. It deserves serious refereeing. I would send it to peer review and ask for a clarification of the efficiency claim, per-dataset seed variance, and a sentence acknowledging that DCFL was previously published. With those changes it is a reasonable journal contribution.","headline":"A genuinely useful oriented tiny-object dataset and benchmark wrapped around a DCFL method that was already published at CVPR 2023; the efficiency claim needs a clarifying sentence but the resource is solid.","tokens_in":30962,"tokens_out":3720,"would_cite":true,"duration_ms":37698,"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 claims that detectors miss 10-pixel objects because training priors and sample selection are biased, and that a dynamic coarse-to-fine training scheme removes the bias, backed by a new dataset and eight-dataset experiments.","keywords":["oriented object detection","tiny object detection","aerial imagery","label assignment","dynamic prior","Gaussian similarity","benchmark dataset","learning bias"],"falsifier":"A direct test would be to compute, on AI-TOD-R, the top-K GJSD candidates for every ground-truth box under 16 pixels and check how often at least one candidate overlaps the box with rotated IoU above 0.5 after the dynamic prior update; if the majority of tiny objects have no such usable candidate, the coarse-to-fine pipeline cannot be supplying the samples the paper credits it with, and the measured gains must come from another mechanism.","tokens_in":29892,"feed_emoji":"🛰️","tokens_out":11018,"duration_ms":95445,"temperature":0.7,"pith_summary":"The paper tries to establish that oriented tiny object detection fails mainly because of a training-time learning bias, not just weak appearance: fixed prior positions rarely land on objects that average 10.6 pixels, and confidence-based sample selection gives those objects too few positive samples, so they are marginalized further. To support this, it introduces AI-TOD-R, an aerial dataset of 752,460 oriented tiny objects, and a benchmark covering fully supervised and label-efficient methods. Its proposed remedy, Dynamic Coarse-to-Fine Learning (DCFL), moves prior positions onto the object's main area with dynamically predicted offsets and splits label assignment into a coarse candidate-selection step followed by a finer per-object quality filter. The paper reports that DCFL improves detection on eight datasets, including the new benchmark, without adding inference-time parameters or computation. A careful reader would care because it suggests the bottleneck is fixable supervision, not receptive fields or backbone capacity.","feed_headline":"Training fix lifts tiny rotated-object accuracy by 16 points","feed_subtitle":"Aerial objects average 10.6 pixels; a dynamic prior and two-stage sample selection stop detectors from ignoring them.","key_machinery":"The load-bearing machinery is the Gaussian parameterization of priors and ground truths combined with two dynamic modules. Each box is represented as a 2-D Gaussian, and the Generalized Jensen-Shannon Divergence (GJSD) between two Gaussians provides a symmetric, scale-invariant similarity that replaces IoU in the candidate-ranking step. The Prior Capturing Block (PCB) generates deformable-convolution offsets that update each prior location by an average offset scaled by feature stride, so priors track the small object's discriminative area instead of sitting fixed on the feature grid. The finer filter is the Dynamic Gaussian Mixture Model (DGMM), a two-component mixture whose means are the object's geometry center and the semantic center implied by the medium-quality samples; samples below a threshold are re-labeled negative. Together these components convert static prior setting and one-shot assignment into a dynamic coarse-to-fine loop.","core_discovery":"The central claim is that across one-stage, two-stage, and query-based detectors, the same bias appears: objects smaller than roughly 16 pixels receive fewer positive training samples than larger objects because prior positions are static and sparse, and they score lower on posterior confidence, which suppresses them under both supervised sample selection and pseudo-label thresholds. The authors argue this bias is a separable cause of poor tiny-object accuracy, and they show it can be measured in the number and quality of assigned samples. Their remedy, DCFL, has two parts: a Prior Capturing Block that predicts per-location offsets and moves each prior toward the object's main area, and a two-stage label assignment in which a coarse positive bag is built using the scale-invariant Generalized Jensen-Shannon Divergence between Gaussian-fitted priors and ground truths, after which a Dynamic Gaussian Mixture Model per instance filters the bag to high-quality samples. The paper reports that this scheme outperforms the compared methods on AI-TOD-R and generalizes to seven further datasets, with the improvement concentrated in tiny classes.","pith_inferences":["Inference (not the paper's claim): if the bias diagnosis is right, then any score-thresholding pseudo-labeling pipeline in semi-supervised detection should under-select tiny objects; replacing the global confidence threshold with a size-calibrated threshold would likely narrow the remaining gap between SSOD and fully supervised training.","Inference: the paper's ablations compare GJSD against KLD and GWD only; a natural next test is to swap in a learned matching network inside the same coarse-to-fine structure, which would reveal how much of the gain is the Gaussian metric versus the dynamic candidate-bag design.","Inference: since the Prior Capturing Block is model-agnostic, the same offset-update idea could be applied to query-based detectors whose priors are already learned, potentially making the dynamic prior component unnecessary there; the paper does not test this configuration."],"forward_implications":["On AI-TOD-R, DCFL raises the AP0.5 of RetinaNet-O from 23.9 to 36.7 at the 1x schedule and of S2A-Net to 49.6 at 40 epochs, without extra inference parameters or FLOPs.","The bias diagnosis extends to label-efficient training: with 30% labeled images, the best semi-supervised method reaches a level comparable to a fully supervised one-stage baseline, and HBB-only supervision approximates OBB supervision, implying annotation cost for oriented tiny objects can be cut sharply.","Applying the same learning scheme to horizontal detectors improves small-object detection on VisDrone, MS COCO, and DOTA-v2 HBB by 2-3 AP0.5 points, so the correction is not specific to rotated boxes.","On multi-scale datasets, per-class gains concentrate on tiny classes, with around 10-18 point improvements on DIOR-R's vehicle, bridge, and wind-mill classes, consistent with the claim that the bias, not the class, was the obstacle.","Ablations on DOTA-v2 val attribute the gain to all three components—dynamic prior, coarse candidate bag, and DGMM filtering—with the full combination reaching 59.15 mAP versus the 51.70 baseline."],"supporting_citations":[{"why":"Supplies the AI-TOD-v2 horizontal-box annotations and tiny-object benchmark that AI-TOD-R inherits and extends with orientation.","marker":"[13]"},{"why":"Weakly supervised H2RBox-v2 generates the initial oriented boxes from horizontal supervision used in AI-TOD-R's semi-automatic annotation.","marker":"[64]"},{"why":"Gaussian receptive field label assignment for tiny objects; basis for modeling priors and ground truths as Gaussians in coarse matching.","marker":"[44]"},{"why":"Gaussian Kullback-Leibler divergence loss for rotated objects; motivates the Gaussian representation of boxes and scale-invariant similarity.","marker":"[70]"},{"why":"Gaussian Wasserstein distance formulation used in rotated object detection; another source of the Gaussian fitting and scale invariance.","marker":"[83]"},{"why":"Defines the Jensen-Shannon divergence used as the starting point for the coarse similarity measure.","marker":"[86]"},{"why":"Provides the closed-form generalized Jensen-Shannon divergence used as the coarse similarity metric between Gaussian priors and ground truths.","marker":"[87]"},{"why":"SODA-A dataset and benchmark supply an independent small oriented object testbed for validating DCFL's generality.","marker":"[10]"},{"why":"Oriented R-CNN is the two-stage architecture into which DCFL is plugged to demonstrate model-agnostic improvement.","marker":"[9]"}],"fun_headline_variants":["DCFL and AI-TOD-R lift tiny oriented object accuracy by 16 points","Dynamic coarse-to-fine learning boosts tiny object detection by 16 points","Tiny rotated objects gain 16 points from unbiased learning scheme","New benchmark and dynamic priors raise tiny-object accuracy 16 points","10.6-pixel objects get 16-point boost from dynamic unbiased training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The scheme assumes that the mathematical similarity used to compare a tiny object with candidate detection boxes keeps ranking the right boxes correctly even when the object spans only about ten pixels, so that the coarse candidate bag and the fine per-object filter both select useful samples; if that similarity stops discriminating at this scale, the reported gains would degrade.","fun_headline_variants_meta":{"raw":{"variants":["DCFL and AI-TOD-R lift tiny oriented object accuracy by 16 points","Dynamic coarse-to-fine learning boosts tiny object detection by 16 points","Tiny rotated objects gain 16 points from unbiased learning scheme","New benchmark and dynamic priors raise tiny-object accuracy 16 points","10.6-pixel objects get 16-point boost from dynamic unbiased training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001307,"raw_usage":{"total_tokens":5355,"prompt_tokens":996,"completion_tokens":4359,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":4262}},"tokens_in":612,"tokens_out":4359,"duration_ms":25413,"temperature":1.0,"reasoning_tokens":4262,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:46:57.861770+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to compute, on AI-TOD-R, the top-K GJSD candidates for every ground-truth box under 16 pixels and check how often at least one candidate overlaps the box with rotated IoU above 0.5 after the dynamic prior update; if the majority of tiny objects have no such usable candidate, the coarse-to-fine pipeline cannot be supplying the samples the paper credits it with, and the measured gains must come from another mechanism.","supporting_citations":[{"cited_title":"H2rbox- v2: Incorporating symmetry for boosting horizontal box supervised oriented object detection,","cited_arxiv_id":null,"evidence_quote":"Weakly supervised H2RBox-v2 generates the initial oriented boxes from horizontal supervision used in AI-TOD-R's semi-automatic annotation."},{"cited_title":"Rfla: Gaussian receptive field based label assignment for tiny object detection,","cited_arxiv_id":null,"evidence_quote":"Gaussian receptive field label assignment for tiny objects; basis for modeling priors and ground truths as Gaussians in coarse matching."},{"cited_title":"Learning high-precision bounding box for rotated object detection via kullback-leibler divergence,","cited_arxiv_id":null,"evidence_quote":"Gaussian Kullback-Leibler divergence loss for rotated objects; motivates the Gaussian representation of boxes and scale-invariant similarity."},{"cited_title":"Rethink- ing rotated object detection with gaussian wasserstein distance loss,","cited_arxiv_id":null,"evidence_quote":"Gaussian Wasserstein distance formulation used in rotated object detection; another source of the Gaussian fitting and scale invariance."},{"cited_title":"A new metric for probability distributions,","cited_arxiv_id":null,"evidence_quote":"Defines the Jensen-Shannon divergence used as the starting point for the coarse similarity measure."},{"cited_title":"On a generalization of the jensen–shannon divergence and the jensen–shannon centroid,","cited_arxiv_id":null,"evidence_quote":"Provides the closed-form generalized Jensen-Shannon divergence used as the coarse similarity metric between Gaussian priors and ground truths."}],"review_version":1}