{"id":"ef50b257-8f48-4f42-9e6b-c542f1b45304","arxiv_id":"2412.05837","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A point-supervised detector with a two-phase denoising teacher-student pipeline achieves state-of-the-art results on tiny object datasets from noisy point annotations.","lead":"Point Teacher is a training method that detects tiny objects in aerial images using only one click per object instead of a full bounding box. It is designed to stay accurate even when the click is off-center, which could cut the cost of labeling large remote-sensing datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Robustness claim rests solely on zero-mean uniform point noise (Eq. 15); systematic annotator bias (e.g., clicks on object heads) would defeat the point-in-box cost in Eq. 8, so real-world comparability to box supervision remains unverified.","rationale":"The reader's weakest_assumption identifies the same load-bearing issue, and I agree with it. The robustness to point-location shifts is the methodological novelty of Point Teacher; if it holds only for zero-mean uniform noise, the practical advantage over box annotation is unestablished. I also weighed the RPN/FCOS inconsistency in §4.1 and the choice of AP0.25, but these are reproducibility and metric issues, not threats to the core mechanism. The paper does provide strong internal support: component ablations in §4.4 show each module contributes, the robustness trend is consistent across datasets, and the DMIL design is clearly specified. None of this resolves the external-validity gap. Since the gap is empirical and addressable with a real-annotation or biased-noise study, the reader's CONDITIONAL verdict remains appropriate; no change is needed.","tokens_in":22010,"tokens_out":5954,"duration_ms":62379,"concrete_test":"Re-run the AI-TOD-v2 robustness experiment (Table 2) with a biased point model: instead of uniform sampling inside the box (Eq. 15), sample each point from a truncated Gaussian centered at the object's head/upper-body with std ≈ 0.3×box size, or with a fixed mean offset of 0.25×w in x. If mAP at the '100% noise' setting drops by more than 10 points relative to the uniform-noise result (31.6), the robustness claim is specific to zero-mean noise. The definitive follow-up is to collect real point annotations on a few hundred AI-TOD-v2 images and compare; the biased simulation is the cheaper first check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central contribution is robustness to point-location noise, but every robustness experiment (Tables 2–3) generates points with Eq. 15: Δx, Δy i.i.d. uniform in [-m/2, m/2], a zero-mean noise model. The denoising machinery is built around this noise. In the point-matching step (Eq. 8), the spatial cost is 0 whenever the annotated point lies inside the predicted box, regardless of where inside; classification score is the only tie-breaker. Under zero-mean uniform noise, boxes centered on the object tend to contain the point and also have high classification scores, so averaging over bags corrects the offset. But human annotators of tiny objects are likely to produce systematic biases—clicks on the visible head of a person, or just outside the faint 16×16 boundary. Under such bias, the set of boxes that contain the point is shifted systematically; the cost matrix in Eq. 8 rewards exactly that shifted set, and no bag construction or jittering can remove a non-zero-mean error because the DMIL refinement (Eq. 10) still seeds from the biased coarse box. The claimed 3.9-point drop at m=100% on AI-TOD-v2 and 7.3-point drop on SODA-A are therefore measured against a distribution that may not match real annotation noise. The paper provides no experiments with human point annotations on any dataset, so the headline 'comparable performance with box-supervised learning methods' is unverified for the setting the method is actually intended for.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Point Teacher, an end-to-end point-supervised method for tiny object detection in aerial images. The method uses a teacher–student architecture with two denoising phases: Spatial-aware Box Generation, which combines random-mask region regression with a Dynamic Multiple Instance Learning (DMIL) module to obtain coarse pseudo boxes, and Noise-aware Label Evolution, which uses point matching, DMIL refinement, and a Jittering IoU loss to generate cleaner pseudo boxes for student supervision. Experiments are reported on AI-TOD-v2 (HBB), SODA-A (OBB), and TinyPerson under synthetic point noise generated by Eq. (15), with mAP gains over prior point-supervised baselines and relatively small degradation when point locations are randomized.","tokens_in":22393,"tokens_out":9097,"duration_ms":83093,"significance":"If the results hold, the paper demonstrates a practical low-cost annotation route for tiny object detection, avoiding auxiliary knowledge such as SAM or synthetic objects. The ablation study (Table 5a) clearly shows that each added component contributes to the final mAP (21.0 → 35.5), and the robustness tables show smaller degradation than comparison methods under the uniform-noise protocol. The method is also general across HBB and OBB detectors. However, the claimed significance is tempered by two issues: the robustness evaluation covers only zero-mean uniform point noise, so the transfer to real human annotation behavior is unverified, and the 'comparable performance with box-supervised learning methods' claim is stronger than the reported numbers support. The paper also does not provide the pseudo-label quality diagnostics that would substantiate the 'progressive denoising' mechanism.","major_comments":[{"comment":"The robustness evaluation is conducted solely under zero-mean uniform point noise, since Δx and Δy are sampled i.i.d. from U(−m/2, m/2) around the box center. Real point annotations for tiny objects are likely to contain systematic biases (e.g., clicks on the visible head or the most salient part). The point-matching cost in Eq. (8) assigns zero spatial cost to any predicted box that contains the annotated point, so a systematic bias would shift the selected boxes toward the biased location, and the DMIL refinement in Eq. (10) cannot remove a non-zero-mean offset because its bags are seeded by the biased coarse boxes. The small degradations in Tables 2–3 therefore validate robustness only under the synthetic uniform-noise model. Please add experiments with human-annotated points or with systematic bias injections (e.g., anisotropic or non-zero-mean noise) to support the claimed robustness for realistic annotation behavior.","section":"§4.1, Eq. (15); Tables 2–3"},{"comment":"The abstract claims that 'relying solely on point supervision, our Point Teacher already shows comparable performance with box-supervised learning methods.' This is not supported by the reported numbers: on AI-TOD-v2 (Table 1) Point Teacher achieves 35.5 mAP vs. 41.2 for FCOS; on SODA-A it achieves 47.2 vs. 74.6 for FCOS-O; and on TinyPerson (Table 4) it reaches 18.6 vs. 34.2 for FCOS, which the authors themselves describe as 54.4% of hbox-supervised accuracy. Moreover, all point-supervised comparisons are reported only at AP0.25, which is a lenient threshold. Please either revise the claim to 'approaching box-supervised performance' or 'substantially outperforming prior point-supervised methods' and discuss the remaining gap, or report results at stricter IoU thresholds to support the comparability claim.","section":"Abstract; Tables 1 and 4"},{"comment":"The paper describes a 'progressive denoising' mechanism, but the teacher network in Phase 2 is an EMA of the student, and the pseudo boxes that supervise the student are generated by this same teacher from the student's own predictions. This creates a self-training loop. Without external clean box labels, the observed mAP gain could stem from better feature learning even if the pseudo boxes do not become more accurate. The paper does not report pseudo-box precision/recall or agreement with GT boxes during training, nor does it compare with a variant where the teacher is frozen or where pseudo boxes are derived purely from point matching. Please add such diagnostics to substantiate the 'denoising' interpretation and to rule out confirmation bias.","section":"§3.1, Eq. (1); Figure 3"},{"comment":"The Spatial-aware Box Generation phase relies on the assumption that predicting randomly masked regions transfers to object-box regression. The only direct evidence is the global ablation in Table 5a, where Phase 1 alone yields 21.0 mAP; this does not isolate the mask-prediction task from the DMIL regression branch that is simultaneously trained on masked regions. Please provide a more direct comparison, e.g., the same architecture with the mask-region regression term removed or with random mask targets, and an analysis of the learned spatial awareness, to support the claimed transfer from mask regression to object-box regression.","section":"§3.2, Eq. (2); Table 5a"}],"minor_comments":[{"comment":"The indexing in Eq. (14) is confusing: the right-hand side uses P_{i-1} both as the accumulated layer and as the iteration variable, and the treatment of i=4 and i in {5,...,7} is inconsistent. Please rewrite with explicit loop indices.","section":"§3.4, Eq. (14)"},{"comment":"The jittering list in Eq. (12) shows only three perturbations followed by '...'; please specify the full set of combinations of width and height perturbations to remove ambiguity.","section":"§3.3, Eq. (12)"},{"comment":"The checkmark layout in Table 5a is not self-explanatory; the column headers and row entries do not align visually. Please reformat to make it clear which components are active in each row.","section":"Table 5a"},{"comment":"It is not stated whether the sampled points are clipped to the image boundaries or to the object box; since m=100% allows points on the box boundary, please clarify the protocol.","section":"§4.1, Eq. (15)"},{"comment":"The term L_cls(s_j^i, c_j) in the cost matrix is not defined in the text; please state which classification loss is used in the point-matching cost.","section":"§3.3, Eq. (8)"},{"comment":"In Table 1, Point2Rbox-RC is listed with E2E = Y, yet the introduction claims that existing point-supervised methods adopt a two-step, non-end-to-end paradigm. Please resolve this inconsistency and qualify the 'first end-to-end' claim to refer specifically to a denoising-based end-to-end framework.","section":"§1, §2.2, Table 1"},{"comment":"The statement that performance decreases by 'only 3.9% and 7.3%' should be worded as '3.9 and 7.3 mAP points' to avoid confusion with relative percentages.","section":"§4.3"},{"comment":"The Discussion's limitation list does not mention the uniform-noise assumption underlying the robustness evaluation; please add an explicit statement that real annotation biases are not modeled and remain as future work.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal. The main revision requirement is to close the gap between the synthetic uniform-noise evaluation and the real annotation scenario; without that, the robustness headline is overclaimed. The 'comparable performance with box-supervised methods' claim should also be moderated or supported with stricter IoU results. The 'first end-to-end' novelty claim should be carefully checked against Point2Rbox (CVPR 2024), which is also end-to-end and point-supervised for oriented aerial detection, even though the benchmarks differ."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Know this: Point Teacher is the first end-to-end point-supervised detector for tiny objects, and on its own terms it works—the ablations show each component earns its keep (21.0 → 35.5 mAP on AI-TOD-v2), and it beats the previous point-supervised baselines by a wide margin. But the two big claims in the abstract—robustness to noisy points and parity with box supervision—rest on shakier ground than the paper lets on.\n\nWhat's genuinely good: the two-phase design is sensible and well-motivated. Random masking gives the model spatial awareness; the DMIL bag extension and refinement is a real improvement over static MIL (Table 5b); the jittering IoU loss is a cheap trick that adds a point. The robustness tables show consistent trends across three datasets under their noise protocol. The authors also state their own limitations clearly in Section 5, which I respect.\n\nThe soft spot is the noise model. All the robustness experiments use points drawn uniformly inside the GT box with zero-mean offsets (Eq. 15). The point-matching cost in Eq. 8 is exactly calibrated to that distribution: the spatial cost is 0 whenever the click is inside the predicted box. So the denoising has a fair shot of averaging out zero-mean shifts. Real human clicks on tiny objects are unlikely to be zero-mean and uniform—think of clicks consistently placed on a car's roof or a person's head, or just outside the boundary because the object is barely visible. Under systematic bias, the set of boxes containing the point is shifted, and the DMIL refinement in Eq. 10 still seeds from the biased coarse box; nothing in the pipeline removes a non-zero-mean error. The paper has no experiments with human point annotations at all. So the robustness claim is only established for one synthetic noise distribution.\n\nSecond, the abstract's 'comparable performance with box-supervised learning methods' is overgenerous. On AI-TOD-v2 with central clicks, Point Teacher gets 35.5 mAP vs FCOS's 41.2 at AP0.25—a 5.7-point gap. That's respectable for point supervision, but it's not parity. On SODA-A, the gap to FCOS-O is even larger (47.2 vs 74.6). And on TinyPerson it reaches 54.4% of box-supervised accuracy. The method is clearly the SOTA among point-supervised methods, but 'comparable' is doing a lot of work.\n\nMinor: the implementation details mention RPN proposals while the main results use FCOS; the paper should settle on one detector. Also no code released yet, which makes the engineering hard to reproduce.\n\nWho should read this: anyone working in weakly supervised detection or remote sensing annotation cost. It deserves a serious referee, but the revision needs to address the noise-model gap—either with human-annotation experiments or with biased-click simulations (head-biased, corner-biased)—and soften the parity claim. As it stands, I'd treat the results as a strong lower bound on what's possible under favorable click noise, not as evidence that real point clicks can replace boxes.","headline":"Good method, honest ablations, but the robustness claim only holds for zero-mean uniform click noise, and the box-parity claim is a stretch.","tokens_in":22917,"tokens_out":4159,"would_cite":true,"duration_ms":34931,"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":"Point clicks alone can supervise tiny-object detection nearly as well as full boxes.","keywords":["point supervision","tiny object detection","aerial images","teacher-student framework","multiple instance learning","denoising","weakly supervised detection"],"falsifier":"Collect a set of human point annotations on AI-TOD-v2 or SODA-A (not generated by Eq. 15), measure the empirical offset distribution, and train Point Teacher on those annotations; if the mAP drops dramatically relative to the reported 31.6–35.5 range, or if deliberately biased clicks (all near the object boundary or outside the box) cause the DMIL point-matching cost to fail, the robustness claim would be falsified.","tokens_in":21842,"feed_emoji":"🎯","tokens_out":4230,"duration_ms":37011,"temperature":0.7,"pith_summary":"This paper argues that a single click per tiny object, instead of a full bounding box, is enough supervision to train a tiny-object detector, and it introduces Point Teacher to make that work. Point Teacher is claimed to be the first end-to-end point-supervised detection framework for tiny objects, and it reports detection accuracy close to fully box-supervised detectors on aerial and person datasets. The method's key move is to treat point labels as noisy and to denoise them in two stages: first the network learns spatial scale from randomly masked image regions, then a teacher-student loop with dynamic multiple-instance learning and a jittered regression loss refines coarse boxes. If the claim holds, tiny-object detection can be labeled with far less annotation effort while retaining most of the accuracy of box-supervised models.","feed_headline":"One click per tiny object rivals full-box supervision","feed_subtitle":"Two-phase denoising pipeline keeps ~35 mAP on aerial tiny objects using point clicks only","key_machinery":"The central object is a two-phase denoising pipeline inside a teacher-student detector. Phase one, Spatial-aware Box Generation, randomly masks image regions and forces the regression head to predict the masked box, giving the network a scale-and-location prior without auxiliary data; Phase two, Noise-aware Label Evolution, matches teacher predictions to annotated points with a two-stage Top-K cost matrix (classification cost plus a term that penalizes boxes not containing the point), then refines the matched pseudo boxes through Dynamic Multiple Instance Learning (DMIL)—constructed proposal bags around each coarse box, extended, scored by classification and instance branches, and fused—and regularizes regression with Jittering IoU Loss, which minimizes against several slightly perturbed versions of the target box. The paper also replaces FPN and standard label assignment with Top-down FPN Aggregation and a scale-invariant one-to-one assignment so that no ground-truth scale information is needed.","core_discovery":"Point Teacher establishes that point-level supervision alone can drive tiny object detection to performance comparable with box supervision: on AI-TOD-v2 it reports mAP 35.5 at AP0.25 with central clicks versus 41.2 for box-supervised FCOS, and under fully randomized point locations it degrades by only 3.9 mAP. The framework is end-to-end and orientation-agnostic (HBB and OBB), and it generalizes across detectors. It works by decoupling learning into Spatial-aware Box Generation, where the teacher learns to turn noisy points into coarse boxes with the help of random mask regression, and Noise-aware Label Evolution, where dynamic multiple instance learning and Jittering IoU Loss refine those boxes into clean pseudo-box supervision for the student. The paper contrasts this with prior MIL, class-probability-map, and auxiliary-knowledge methods, which collapse when point clicks shift.","pith_inferences":["Editorial inference: If human annotators produce biased point distributions rather than uniform random offsets, the synthetic noise model may underestimate the difficulty; a human-in-the-loop study or annotation-bias analysis would clarify.","Editorial inference: The denoising approach could combine with semi-supervised or active learning, using point clicks as an inexpensive query primitive for tiny objects across large aerial scenes.","Editorial inference: The random-mask spatial awareness trick might extend beyond point supervision to other sparse or noisy label settings, such as weak labels in remote sensing."],"forward_implications":["Point supervision is a viable low-cost substitute for box supervision on tiny objects, cutting annotation effort to roughly one click per object.","The two-phase denoising paradigm transfers to both horizontal and oriented bounding box detectors without architectural changes.","Existing point-supervised methods (MIL, CPM, and auxiliary-based) are insufficiently robust to off-center clicks for tiny objects.","The performance gap to box-supervised models is modest and the method retains most accuracy even at 100% randomization of click positions."],"supporting_citations":[{"why":"Supplies the neighbor-sampling bag construction strategy and the MIL baseline that Point Teacher compares against and outperforms.","marker":"(Chen et al., 2022)"},{"why":"Supplies the EMA teacher-student pseudo-labeling pipeline that Point Teacher adapts with point matching.","marker":"(Xu et al., 2021b)"},{"why":"Motivates the random-masking self-supervised objective used for Spatial-aware Box Generation.","marker":"(Oquab et al., 2023)"},{"why":"Serves as the base detector architecture used in the main horizontal bounding box experiments.","marker":"(Tian et al., 2019)"},{"why":"Supplies the RoIAlign/ROI extractor used by the DMIL module.","marker":"(Ren et al., 2017)"},{"why":"Provides the strongest oriented bounding box point-supervised baseline on SODA-A that Point Teacher compares against.","marker":"(Ren et al., 2024)"},{"why":"Provides the CPM-based baseline on AI-TOD-v2 that Point Teacher compares against.","marker":"(He et al., 2024)"}],"fun_headline_variants":["Single point per object rivals full boxes in tiny-object detection","Point clicks outperform expectation: 35.5 mAP on tiny objects","One click per tiny object: box-level accuracy from point-only labels","Point Teacher turns noisy clicks into robust tiny-object detection","Point-only supervision nearly matches box supervision on tiny objects"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The main load-bearing premise is that real point annotations for tiny objects behave like points drawn uniformly at random inside the ground-truth box with a controllable maximum offset; all robustness experiments use this synthetic noise model, so systematically biased clicks (for example, always on a person's head or just outside the object) could break the pipeline.","fun_headline_variants_meta":{"raw":{"variants":["Single point per object rivals full boxes in tiny-object detection","Point clicks outperform expectation: 35.5 mAP on tiny objects","One click per tiny object: box-level accuracy from point-only labels","Point Teacher turns noisy clicks into robust tiny-object detection","Point-only supervision nearly matches box supervision on tiny objects"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1571,"prompt_tokens":979,"completion_tokens":592,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":508}},"tokens_in":595,"tokens_out":592,"duration_ms":6538,"temperature":1.0,"reasoning_tokens":508,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:17:08.688988+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a set of human point annotations on AI-TOD-v2 or SODA-A (not generated by Eq. 15), measure the empirical offset distribution, and train Point Teacher on those annotations; if the mAP drops dramatically relative to the reported 31.6–35.5 range, or if deliberately biased clicks (all near the object boundary or outside the box) cause the DMIL point-matching cost to fail, the robustness claim would be falsified.","supporting_citations":[],"review_version":1}