{"id":"929a62dc-fe66-48db-986e-73f835c37001","arxiv_id":"2507.22438","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A domain adaptation method uses event-derived motion to synthesize blur and iteratively cleans pseudo-labels, improving multi-person 2D pose estimation on blurry images without annotations.","lead":"This paper adapts 2D human pose estimation models to blurry images without target-domain annotations by using event cameras to create motion-aware blur for training. It also refines pseudo-labels through a teacher-student scheme, reporting gains over prior domain adaptation methods on the EHPT-XC dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The direct evidence for the central event-based blur augmentation (Table 3) is internally inconsistent and shows only +0.2–0.4 mAP; without a random-blur control or seed statistics, the claim that motion-aware augmentation bridges the domain gap is unsupported.","rationale":"The reader's weakest assumption—that the synthetic event-based blur is a faithful proxy for real motion blur—is relevant, but the more immediate and load-bearing issue is evidentiary: the paper's own isolated measurement of the augmentation is internally inconsistent and lacks the control needed to attribute the effect to event-derived motion. The central claim names event-based augmentation as the mechanism that 'effectively bridges' the domain gap, so if that mechanism is not measurably better than random blur, the paper's headline contribution is not established. This does not overturn the overall conditional verdict: the pipeline still shows gains, and the issue is fixable with a proper ablation and corrected reporting. It does sharpen the condition: accept only after the augmentation is shown to be both statistically reliable and specifically motion-aware, not just any blur augmentation. I therefore keep the reader's CONDITIONAL verdict rather than moving to reject, and I mark agreement as partial because the reader focused on the fidelity assumption but did not flag the Table 3 text/table mismatch or the absence of the random-blur control.","tokens_in":16144,"tokens_out":5603,"duration_ms":65527,"concrete_test":"Re-run Stage 1 with at least 3 seeds for: (a) baseline sharp-only training, (b) event-based motion augmentation as in Sec. 3.2, (c) random motion-blur augmentation using randomly sampled kernels/directions, and (d) event augmentation plus the Blur2Blur translation loss. Report mean ± std of target mAP/mAR. The central claim requires (b) to exceed (a) by more than 2× the pooled seed std and (b) to exceed (c) by a similar margin. Additionally, correct Table 3 and its accompanying text so the numbers agree; currently the '+1.8' claimed improvement is a mAR change, not mAP, and the mAP deltas are 0.2 and 0.4.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that event-based motion-aware augmentation bridges the sharp-to-blur domain gap (Sec. 3.2, Eq. 2). The only isolated test of this mechanism is Table 3, but the table and the text disagree. The text says 'motion augmentation improves mAP by 0.4' and that combined with blur translation it achieves 'a total improvement of 1.8'; however, the table shows Baseline 46.5, +Motion Aug. 46.7, and +Motion Aug. +Blur Translation 46.9. That is +0.2/+0.4 mAP, and the '+1.8' figure corresponds to mAR (50.4→52.2), not mAP. So the headline mechanism's mAP contribution is at most 0.4, potentially within seed noise. More fundamentally, there is no control that trains the Stage-1 teacher with random blur kernels, so the observed gain cannot be attributed to event-derived motion rather than to 'any realistic blur augmentation'. The later stages add further gains, but those gains do not rescue the specific claim that event-based motion awareness is what bridges the domain gap. If the motion-aware augmentation is not measurably better than random blur, the central novelty—and the 'event cameras as bridge' narrative—is unsupported, even though the overall pipeline may still be useful.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an unsupervised domain adaptation method for multi-person 2D pose estimation from sharp images to motion-blurred images, using event cameras as an auxiliary modality. In Stage 1, a multi-modal teacher network (image+event and event-only sub-networks fused by a refinement module) is trained on source-domain sharp data augmented with synthetic motion blur generated by forward-warping sharp frames along event-derived optical flow, optionally made more realistic by a blur translation loss. In Stage 2, the teacher produces pseudo-labels for the unlabeled target blur domain, and a student network is trained with confidence-based masking. Stages 3 and 4 iteratively refine the teacher and student using a mutual uncertainty masking scheme that suppresses inconsistent predictions across the two networks. Experiments on the EHPT-XC dataset show that the full pipeline outperforms existing UDA pose-estimation baselines and approaches, but the isolated contribution of the event-based blur augmentation (Table 3) is small. The paper includes an ablation study, threshold analysis, and an image-only testing variant, and it releases code.","tokens_in":16458,"tokens_out":5461,"duration_ms":60779,"significance":"If the central claims hold, this is a novel application of event cameras to unsupervised domain adaptation for pose estimation under motion blur, and it demonstrates a practically useful result: the full pipeline improves target-domain mAP from 37.5 to 51.6 over the no-adaptation baseline and outperforms prior UDA methods. The paper's strength is the comprehensive system design and the use of a real benchmark (EHPT-XC) with synchronized image/event data. The authors also provide qualitative comparisons and an image-only student variant, which increases practical relevance. However, the significance is currently tempered by weak direct evidence for the headline mechanism: the motion-aware blur augmentation alone yields only 0.2–0.4 mAP improvement over a no-augmentation teacher baseline, with no random-blur control or seed-variance reporting, and the paper overstates its oracle-comparability. The contribution is potentially important, but the quantitative support for the core novelty needs substantial strengthening.","major_comments":[{"comment":"The text in 'Effectiveness of the motion-aware augmentations' states that 'motion augmentation improves mAP by 0.4, and when combined with blur translation, it achieves a total improvement of 1.8.' This is inconsistent with Table 3, which reports mAP of 46.5, 46.7, and 46.9 for Baseline, +Motion Aug., and +Motion Aug. + Blur Translation, respectively (i.e., +0.2 and +0.4 mAP); the +1.8 figure corresponds to mAR (50.4→52.2), not mAP. The central novelty's mAP contribution is therefore at most 0.4, which could be within seed noise. More importantly, there is no ablation that trains the Stage-1 teacher with randomly generated blur kernels, so the evidence does not establish that event-derived motion, rather than generic blur augmentation, is what bridges the domain gap. Please add a random-blur control, report results over multiple seeds (or seed variance), and correct the metric misstatement in the text.","section":"Sec. 5, Table 3; Sec. 3.2"},{"comment":"The final method uses masking thresholds th=0.1 and th'=0.1, and Table 6 reports performance under different threshold values. The paper does not state how these thresholds were selected. If they were chosen by sweeping values and picking the ones that maximize target-domain mAP on the test set (or any labeled target data), then the reported final numbers are not obtained in a purely unsupervised manner and the gain from the pseudo-label refinement stages could be inflated. Please specify the threshold selection procedure explicitly; if a held-out labeled validation split was used, describe how it was created; otherwise, fix the thresholds a priori and present the sensitivity analysis without selecting on the test set.","section":"Sec. 5, Table 6; Sec. 4.2"},{"comment":"The claim that the proposed method 'achieves performance comparable to the multi-modal oracle' is not supported by Table 1. The final student network reaches 51.6 mAP on the Target (Blur) set, while the multi-modal oracle Base (I+E) [16] reaches 58.8 mAP, a gap of 7.2 mAP. This is a substantial difference, not 'comparable' in the usual sense. The same overstatement appears in the abstract and conclusion. Please revise the claim to a more precise statement (e.g., 'outperforms existing UDA baselines and reduces the gap to the oracle') so that the paper's contribution is reported accurately.","section":"Sec. 4.3, Table 1; Abstract; Sec. 6"}],"minor_comments":[{"comment":"The summation notation in Eq. (2) is difficult to read; the limits and the meaning of δt(x,y) should be typeset more clearly, and the masked-average operation should be described in words as well as equations.","section":"Sec. 3.2, Eq. (2)"},{"comment":"The overall framework figure is very dense and hard to parse, especially the arrows and label flow between stages; consider simplifying the diagram and using clearer separation between the four stages.","section":"Fig. 2"},{"comment":"The event-based optical flow network is trained with a self-supervised loss, but no architectural details or training specifics are provided beyond references [13,21,54]; a brief description or a table of hyperparameters would help reproducibility.","section":"Sec. 3.2"},{"comment":"The sentence 'Since DualTeacher [1]’s original augmentation is tailored for low-light conditions, we substituted it with standard augmentations' is grammatically incomplete and should be rephrased.","section":"Sec. 4.3"},{"comment":"The term 'Blur Translation' is used in Table 3 but not defined at first use; clarify that it refers to the blur translation loss of [55] and explain how it is applied to the generated blurry images.","section":"Table 3 and Sec. 5"},{"comment":"The row 'Ours-Teacher' is not explained in the text; clarify whether it is the teacher network after Stage 1, after Stage 3, or after all refinement stages.","section":"Table 1"},{"comment":"There is a typo in the figure caption: 'groud-truth' should be 'ground-truth'.","section":"Fig. 2 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant problem and the full pipeline achieves strong results on EHPT-XC, and the code release is a positive. However, the central claim about event-based motion augmentation is supported only by a weak and internally inconsistent ablation (Table 3), and the threshold selection procedure is unclear. These are fixable with additional experiments and clarification, so I recommend major revision rather than rejection. I would urge the authors to add a random-blur control and multi-seed statistics, and to correct the metric misreporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a real contribution: the first UDA method that uses event cameras to adapt pose estimation from sharp to motion-blurred images, and the full pipeline beats prior UDA baselines by a wide margin on the EHPT-XC benchmark. Second, the paper's headline mechanism—event-based motion-aware blur augmentation—does not hold up under scrutiny: in isolation it buys about 0.4 mAP, and the paper lacks a random-blur control, so it is not shown to be the reason the method works.\n\nThe good parts. The problem setup is new and well motivated. The event-based augmentation idea—forward-warping sharp frames along event-derived optical flow to synthesize blur—is sensible and the implementation is reasonable. The mutual uncertainty masking is a decent co-training variant. The final result of 51.6 mAP on target blur versus 34.6 for the best prior UDA method is a real improvement, and the authors release code.\n\nThe soft spots. Table 3 is the only isolated test of the claimed bridging mechanism, and the text misreports it: the table shows +0.2 mAP for motion augmentation and +0.4 with blur translation, while the text claims 0.4 and then 1.8—the 1.8 is actually the mAR gain. More importantly, there is no control that replaces event flow with random blur kernels. Without that, the gain could be just \"any realistic blur augmentation.\" The later stages (pseudo-label refinement) contribute the larger gains (49.1 to 51.6), but those are not specific to events either. Also, the \"comparable to oracle\" claim is overstated: the oracle I+E network gets 58.8 target mAP, ours 51.6.\n\nOne more issue: the thresholds in Stages 2 and 4 are tuned on the target test set, which can inflate results. That is common but still worth flagging.\n\nNone of this invalidates the overall pipeline. The final numbers are good and the adaptation framework is sensible. But the central scientific claim—that event-based motion awareness is what bridges the gap—is not yet supported. A revision with a random-blur control, corrected numbers, and softer claims would make a solid paper.\n\nWho is this for? Researchers working on pose estimation under blur or event-camera perception. It deserves a serious referee: the idea is novel, the benchmark is useful, and the pipeline works. I would send it to review with a request for the control experiment.","headline":"A useful event-camera UDA pipeline for pose under blur, but the headline augmentation's isolated gain is ~0.4 mAP and lacks a random-blur control; the final results are strong enough to warrant review.","tokens_in":16985,"tokens_out":2198,"would_cite":true,"duration_ms":23153,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Event cameras teach pose models to see through blur","keywords":["unsupervised domain adaptation","2D human pose estimation","event cameras","motion blur","event-based blur augmentation","pseudo-label refinement","student-teacher framework","multi-modal fusion"],"falsifier":"Retrain the Stage-1 teacher with the event-derived flow in the augmentation replaced by random flow of the same magnitude; if target-domain mAP does not drop below the true-event-flow model, the motion information is not doing the work. Additionally, compute an image distribution distance such as FID between the synthesized blurry images and real blurry target frames from EHPT-XC; a large distance would indicate the synthetic blur does not actually approximate real blur.","tokens_in":15949,"feed_emoji":"⚡","tokens_out":11450,"duration_ms":106899,"temperature":0.7,"pith_summary":"This paper argues that event cameras can close the domain gap between sharp and motion-blurred images for 2D human pose estimation, without needing any pose annotations in the blurred target domain. Because event streams record per-pixel brightness changes at microsecond resolution, they encode exactly the motion that creates blur; the paper converts that motion into synthetic blur by warping sharp frames along event-derived optical flow and averaging the warps. A multi-modal teacher network trained on this synthetic blur generates pseudo-labels on real blurred frames, and a student–teacher loop with mutual uncertainty masking refines those labels by trusting each network where the other agrees. If the claim holds, pose models trained on ordinary sharp-labeled data plus a synchronized event stream would remain accurate during fast motion and low-light capture, a regime where standard supervision is difficult and expensive to obtain.","feed_headline":"Event cameras teach pose models to see through blur","feed_subtitle":"Sharp-labeled frames plus a synchronized event stream close the blur domain gap, no target annotations needed.","key_machinery":"Two mechanisms carry the argument. The first is motion-aware event-based blur augmentation: the event stream is sliced into time segments, a self-supervised event optical-flow network estimates a flow for each slice, and the sharp source image is forward-warped along each flow and combined with a masked average (Eq. 2) to synthesize a blurry image that reflects true pixel motion; a blur translation loss then pushes the synthesized blur closer to real continuous blur. The second is mutual uncertainty masking within a student–teacher loop: confidence scores for a proposed pose are computed from both the teacher's and the student's heatmaps, and any region where the two disagree below a threshold is excluded from the pseudo-label loss, preventing low-quality labels from driving adaptation. The teacher itself is multi-modal, combining an image+event sub-network and an event-only sub-network through a channel-attention refinement module, so dense spatial evidence and motion-invariant edge evidence are balanced when pseudo-labels are generated.","core_discovery":"The central claim is that motion information from event cameras is sufficient to bridge the sharp-to-blur gap in multi-person 2D pose estimation under fully unsupervised target-domain adaptation. The paper shows that by slicing an event stream into short intervals around a sharp frame's exposure center, estimating optical flow for each slice with a self-supervised network, forward-warping the sharp image along each flow, and masked-averaging the warps (Eq. 2), a model can be trained on synthetic blur that behaves like real blur; a blur translation loss further reduces the residual gap. The resulting teacher, which fuses an image+event sub-network, an event-only sub-network, and a channel-attention refinement module, produces pseudo-labels that a student learns from, with mutual uncertainty masking suppressing disagreeing and thus unreliable predictions. On the EHPT-XC benchmark, the final student reaches 51.6 mAP on the blurred target split, versus 36.4 for the best prior multi-modal domain-adaptation method and 58.8 for an image+event oracle trained with target blur labels, which the paper reads as evidence that event-based augmentation plus mutual masking can substitute for target-domain annotations.","pith_inferences":["The same event-flow warping could synthesize other motion-dependent degradations beyond blur, such as rolling-shutter distortion or motion-based aliasing, giving a general recipe for turning events into training signal for any motion-sensitive perception task.","If events are available only once during adaptation rather than at deployment, the image-only student result hints at a practical deployment pattern: capture events and sharp frames together, adapt, then ship an image-only model.","The mutual-masking logic is modality-agnostic; it should generalize to any sensor pair where one modality is dense but vulnerable to degradation and the other is sparse but robust, such as thermal+RGB or depth+RGB, as a way to generate cleaner pseudo-labels."],"forward_implications":["Pose estimation can be adapted to extreme motion blur using only sharp images with labels plus synchronized event streams, removing the need for paired sharp–blur capture rigs or manual blur annotation.","The same pretrained teacher can supervise an image-only student, so at deployment an ordinary RGB camera suffices; the paper reports 46.0 mAP for the image-only student versus 36.1 for the image-only sharp-trained baseline.","The mutual uncertainty masking makes pseudo-label quality less sensitive to the confidence threshold than single-teacher alternatives, as shown by the threshold analysis, which suggests the method transfers to other unlabeled target domains without per-domain threshold tuning.","Because the augmentation is driven by actual motion rather than random kernels, the domain bridge should hold for object motion of varied speed and direction, not just a fixed blur pattern."],"supporting_citations":[{"why":"Supplies the EHPT-XC dataset with aligned sharp, blur, and event streams plus pose annotations, defining the source/target split and the image+event fusion network the teacher is built on.","marker":"[16]"},{"why":"Provides the bottom-up pose network (center heatmap, offsets, keypoint heatmaps) used as the backbone for teacher, student, and baselines.","marker":"[24]"},{"why":"Provides the self-supervised event optical flow learning that supplies the pixel motion used for warping in the blur augmentation.","marker":"[13]"},{"why":"Contrast maximization framework used as one of the self-supervised losses for the event flow network in the augmentation.","marker":"[21]"},{"why":"Photometric-constancy self-supervised event-to-image reconstruction used to train the event flow network.","marker":"[54]"},{"why":"Blur2Blur blur translation loss that pushes the discretely warped synthetic blur toward real continuous blur, closing part of the residual domain gap.","marker":"[55]"},{"why":"DualTeacher is the main domain-adaptive pose estimation baseline that the method is compared against.","marker":"[1]"},{"why":"UDA-HE is the unified domain-adaptive pose estimation baseline whose performance the method exceeds on the blurred target split.","marker":"[35]"}],"fun_headline_variants":["Event cameras bridge sharp-to-blur pose gap","Unsupervised pose adaptation via event-based blur","Event streams train pose models to beat motion blur","Sharp-to-blur pose adaptation without target labels","Event-based motion data closes blur pose gap"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that blur synthesized by forward-warping sharp images along event-derived optical flow is a faithful enough stand-in for real motion blur in the target domain, and that synchronized event streams will be available wherever the adapted model is deployed.","fun_headline_variants_meta":{"raw":{"variants":["Event cameras bridge sharp-to-blur pose gap","Unsupervised pose adaptation via event-based blur","Event streams train pose models to beat motion blur","Sharp-to-blur pose adaptation without target labels","Event-based motion data closes blur pose gap"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000257,"raw_usage":{"total_tokens":1602,"prompt_tokens":995,"completion_tokens":607,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":537}},"tokens_in":611,"tokens_out":607,"duration_ms":6783,"temperature":1.0,"reasoning_tokens":537,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:39:37.812782+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the Stage-1 teacher with the event-derived flow in the augmentation replaced by random flow of the same magnitude; if target-domain mAP does not drop below the true-event-flow model, the motion information is not doing the work. Additionally, compute an image distribution distance such as FID between the synthesized blurry images and real blurry target frames from EHPT-XC; a large distance would indicate the synthetic blur does not actually approximate real blur.","supporting_citations":[{"cited_title":"A benchmark dataset for event-guided human pose estimation and tracking in extreme conditions","cited_arxiv_id":null,"evidence_quote":"Supplies the EHPT-XC dataset with aligned sharp, blur, and event streams plus pose annotations, defining the source/target split and the image+event fusion network the teacher is built on."},{"cited_title":"Bottom-up human pose estimation via disentan- gled keypoint regression","cited_arxiv_id":null,"evidence_quote":"Provides the bottom-up pose network (center heatmap, offsets, keypoint heatmaps) used as the backbone for teacher, student, and baselines."},{"cited_title":"Temporal event stereo via joint learning with stereoscopic flow","cited_arxiv_id":null,"evidence_quote":"Provides the self-supervised event optical flow learning that supplies the pixel motion used for warping in the blur augmentation."},{"cited_title":"A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estima- tion","cited_arxiv_id":null,"evidence_quote":"Contrast maximization framework used as one of the self-supervised losses for the event flow network in the augmentation."},{"cited_title":"Blur2blur: Blur conversion for unsupervised image deblurring on unknown domains","cited_arxiv_id":null,"evidence_quote":"Blur2Blur blur translation loss that pushes the discretely warped synthetic blur toward real continuous blur, closing part of the residual domain gap."},{"cited_title":"A unified framework for domain adaptive pose estimation","cited_arxiv_id":null,"evidence_quote":"UDA-HE is the unified domain-adaptive pose estimation baseline whose performance the method exceeds on the blurred target split."}],"review_version":1}