{"id":"2d2d3b00-6324-4336-abc6-4f129f95bd93","arxiv_id":"2505.09376","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"AfforDance is a proposed AR system that converts dance videos into interactive lessons using 3D avatars, audio cues, and body-size-adapted visual affordances.","lead":"This paper introduces AfforDance, an augmented reality system that turns a user-selected dance video into an interactive lesson with a 3D avatar and visual overlays on hands and feet. It is a design proposal that may interest HCI researchers working on personalized, screen-based dance instruction.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"AfforDance's learning claim rests on unvalidated affordance alignment: Section 3.2.2 applies normalization only initially and 'confirms' the transform without measurement, so the core visual-guidance mechanism may mislead users when they move.","rationale":"The reader's weakest assumption is essentially the same as the load-bearing concern identified here: the untested accuracy of the body-size-normalized affordance overlay. My stress-test sharpens this by locating the technical fragility in Section 3.2.2 (one-time normalization, 'confirmed' transform, no reported pose-estimation accuracy) and by tying it to the authors' own Section 4 limitation about passive content delivery. There is no internal contradiction that would make the system incoherent; the pipeline is plausible and the CI, while small, is a legitimate design-study input for a workshop paper. However, the abstract explicitly claims that the system 'enhances learning,' and that claim requires at minimum that the affordances remain aligned with the learner's moving body. Neither a user study nor a quantitative alignment validation is provided. This does not justify rejection, because the contribution could reasonably be reframed as a design proposal with future evaluation. It also does not justify a stronger accept, because the flagship assertion is unsupported. Therefore the existing CONDITIONAL verdict is appropriate; the paper should be accepted only if the learning-enhancement claim is softened or accompanied by the proposed alignment or learning-benefit evaluation. I see no reason to move the verdict.","tokens_in":5321,"tokens_out":3355,"duration_ms":38089,"concrete_test":"Run a controlled alignment test: have a participant perform an 8-count dance segment in front of the same webcam setup used by AfforDance, with ground-truth wrist and ankle positions recorded by an optical motion-capture system or by manually annotated 2D points. Then, for each frame, compute the 2D/3D distance between the displayed affordance overlay locations and the tracked wrist/ankle joints. Report median and 90th-percentile error over the segment, and compare the 'initial normalization only' version against a version with per-frame updates. If the median error exceeds about 10 cm or 5% of the user's height, the affordance alignment is too inaccurate to support the claimed learning benefit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that AfforDance 'enhances learning through visual affordances' depends on the affordances being spatially aligned with the user's body while dancing. This alignment is the enabling condition for every downstream learning benefit, yet it is not established. Section 3.2.2 states that normalization is performed once, 'rather than every frame,' and that 'the transform is confirmed to ensure proper alignment.' No quantitative error metric is reported. There are two possible readings, and both are problematic. First, if real-time pose estimation is used only for initial body-size normalization, then as the user moves around the webcam view, the fixed affine transform will drift: translations, rotations, and scale changes due to distance are not corrected, so hand and foot affordances will no longer sit on the corresponding body parts. Second, if ThreeDPoseUnityBarracuda is used per-frame for continuous overlay, the paper never reports its joint localization accuracy under dance-like motion, where fast movements, self-occlusion, and clothing variability are common. Either way, the visual cues that are supposed to guide movement could instead point to the wrong locations, undermining the learning mechanism rather than supporting it. The authors' own Section 4 concedes that user input is not effectively integrated beyond passive content delivery, which further weakens the idea that the system provides reliable, adaptive real-time guidance. The contextual inquiry and the system pipeline are reasonable design contributions, but the abstract's 'enhances learning' claim is not supported by the current evidence or by the alignment validation in the implementation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"AfforDance is an AR-based dance learning system that converts a user-selected dance video into a structured learning experience. The pipeline extracts audio, adds an 8-count beat, generates a 3D reference avatar using the WHAM pose-estimation model, and renders real-time visual 'affordances' overlaid on the user's webcam view using ThreeDPoseUnityBarracuda for initial body-size normalization. Design goals are derived from a contextual inquiry with three participants. The paper presents the system architecture and interface but reports no user study, no quantitative evaluation, and no measured performance data. Section 4 acknowledges that user input is not effectively integrated beyond passive content delivery and identifies this as a key limitation.","tokens_in":5641,"tokens_out":2863,"duration_ms":31102,"significance":"The problem addressed is relevant: self-directed dance learning with AR overlays for large displays is a reasonable alternative to HMD-based approaches, and the use of WHAM for avatar generation is a sensible technical choice. The paper's strengths include a clear system pipeline, the grounding of design goals in a small contextual inquiry, and an honest statement of limitations. However, the central claim that the system 'enhances learning through visual affordances' is entirely unsupported by evidence; the paper is a system description without any empirical validation. The alignment mechanism, which is the enabling condition for the visual guidance, is described as a one-time normalization with no error measurement. If the result holds, the system could be a useful contribution to personalized dance education, but as presented the evidence is insufficient to support the learning claim.","major_comments":[{"comment":"The affordance alignment mechanism is load-bearing but unvalidated. The text states that 'normalization is performed initially rather than every frame' and that 'the transform is confirmed to ensure proper alignment,' without reporting any quantitative error metric. If the normalization is a fixed affine transform, any translation, rotation, or scale change of the user relative to the webcam during dancing will cause the affordances to drift off the corresponding body parts. If ThreeDPoseUnityBarracuda is used per-frame, the paper does not report its keypoint accuracy under dance-like motion with fast movements and self-occlusion. In either reading, the visual cues could misguide rather than guide the learner, undermining the abstract's claim that the system enhances learning through visual affordances. The authors should provide a quantitative evaluation of alignment error during representative dance motions, or describe a continuous re-normalization scheme.","section":"Section 3.2.2"},{"comment":"The paper claims that AfforDance 'enhances learning through visual affordances,' but no user study or learning-outcome evaluation is reported. The only empirical input is a contextual inquiry with three participants that motivates design goals, not an assessment of the implemented system. Given that the central contribution is a claimed learning benefit, the absence of even a small pilot study (e.g., movement accuracy, completion time, or user-reported experience compared with a video-only baseline) leaves the main claim unsupported. Section 4's acknowledgment that 'user input can be effectively integrated into the learning process beyond passive content delivery' further weakens the adaptive-guidance aspect. The manuscript should either add a pilot evaluation or carefully scope its claims to system design and feasibility.","section":"Abstract and Section 4"},{"comment":"The audio content generation and reference avatar generation are described but not validated. For audio, the procedure of adjusting BPM and zero-padding for alignment could introduce temporal misalignment between the beat and the dance video; no measure of synchronization accuracy is provided. For the avatar, WHAM is used to estimate 3D poses from 2D dance videos, but the paper does not address how pose-estimation errors under fast, complex choreography affect the quality of the reference avatar and the subsequent affordances. These pipeline stages are central to the learnability of the content, so some form of accuracy assessment or at least a qualitative example with error overlay would strengthen the contribution.","section":"Section 3.1"}],"minor_comments":[{"comment":"The sentence 'It is hard to watch myself in the mirror and the device at the same time' is presented as a quote from P1 and P3, but it is unclear whether this is a direct quote or a paraphrase; please clarify or use quotation marks consistently.","section":"Section 2.2.1"},{"comment":"The statement 'the transform is confirmed to ensure proper alignment' is vague; specify whether this is visual inspection, manual calibration, or an automatic check.","section":"Section 3.2.2"},{"comment":"The use of 'PKL' and 'FBX' file formats may be unfamiliar to readers outside graphics; a one-sentence explanation would improve accessibility.","section":"Section 3.2.2"},{"comment":"The caption of Figure 3 describes the three affordance display modes, but the text does not explain why a user would choose one mode over another; a brief rationale would help.","section":"Figure 3"},{"comment":"The future direction paragraph is well-articulated, but the phrasing 'key limitation of this study is the insufficient discussion' undersells the issue; it is not only insufficient discussion but also an unvalidated system.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"This manuscript appears to be a workshop submission rather than a full archival paper, given its length and the absence of evaluation. The technical approach is plausible, but the central learning claim is not supported. As a full paper in this journal, it would benefit from at least a small user study with quantitative measures. If the venue expects work-in-progress contributions, the authors should adjust their claims accordingly. The alignment concern in Section 3.2.2 is the most substantive technical risk and should be addressed directly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a workshop paper describing an AR dance learning system that turns any YouTube dance video into a learning session with a 3D avatar, beat track, and body-size-normalized visual cues on hands and feet. The genuinely new bit is the integration: user-selected video to interactive AR overlay on a large screen, avoiding HMDs. That fills a real niche, and the contextual inquiry with three participants gives some user grounding for the design goals. The paper is clearly written and honest about its own limitations.\n\nWhat it does well: the pipeline is coherent—WHAM for offline 3D pose, ThreeDPoseUnityBarracuda for real-time tracking, Unity for rendering—and the interface features (speed control, repeat, section navigation) are sensible for self-learners. The authors cite relevant prior work (YouMove, DancÆR, VR feedback) and explicitly acknowledge in Section 4 that user input is not effectively integrated beyond passive content delivery.\n\nWhere it falls short: the abstract claims the system 'enhances learning through visual affordances,' but there is no user study, no quantitative error metric, and no data at all. The stress-test point is fair: Section 3.2.2 says normalization is done once, not per frame, and the transform is 'confirmed' rather than validated. If the user moves, the fixed affine transform will drift, so the affordances could sit on the wrong places and mislead rather than guide. That is a load-bearing assumption because the entire learning mechanism depends on alignment. The paper also does not report the real-time pose estimator's accuracy under dance-like motion.\n\nThat said, the problem is proportionate to the venue. This is a workshop design contribution; the claims should just be scoped accordingly. A pilot study with a few participants measuring alignment error or learning outcome would go a long way.\n\nWho it's for: people working on AR/VR motor skill learning, dance interfaces, and pose-driven feedback. They will get a clean example of a video-to-AR pipeline and a testable set of design decisions. The paper deserves a serious referee—the system is plausible and the writing is careful—but I would push for either a small evaluation or a softened abstract before acceptance.","headline":"A clearly written workshop system proposal whose 'enhances learning' claim is not backed by any evaluation; the alignment step that everything depends on is 'confirmed' rather than measured.","tokens_in":6138,"tokens_out":2240,"would_cite":false,"duration_ms":23097,"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":"AfforDance turns any user-selected dance video into an AR lesson with body-matched visual cues.","keywords":["augmented reality","dance learning","visual affordance","3D pose estimation","8-count beat","self-paced learning","large display","personalized learning"],"falsifier":"Record the learner's webcam view during a practice session and compare the projected positions of the affordance overlays against the tracked wrist and ankle positions frame by frame; if the average alignment error grows with movement speed or dance range, the one-time normalization is not sufficient. A controlled comparison of learners practicing an 8-count phrase with affordances versus with a plain mirrored video, measuring joint-angle error against the reference pose, would settle whether the cues improve learning.","tokens_in":5068,"feed_emoji":"💃","tokens_out":4968,"duration_ms":48616,"temperature":0.7,"pith_summary":"AfforDance is an augmented-reality system for learning dance from ordinary videos. The paper argues that learners who want a specific choreography should not have to wait for instructor-made lessons or wear a head-mounted display; instead, the system takes any dance video the user selects, extracts a 3D reference avatar, adds an 8-count beat track, and overlays body-size-normalized visual cues on the learner's own webcam image. Those cues, called affordances, highlight the hands and feet where movement is hardest to see, and the interface offers speed control, repetition, and section navigation so learners set their own pace. If the system works as intended, self-directed dance practice becomes more like a mirror-based class without fixed times, places, or instructors.","feed_headline":"AR system turns any dance video into a personalized lesson","feed_subtitle":"Visual cues overlaid on your webcam view show hand and foot moves synced to an 8-count beat.","key_machinery":"The load-bearing object is the visual affordance: an AR overlay, sized to the learner's body, that marks the hands and feet (with optional full-body display) so joint positions and movement ranges are explicit during practice. It is produced by separating vertex groups from the WHAM-generated avatar mesh in Blender and then normalized to the learner's webcam frame using real-time pose estimation from ThreeDPoseUnityBarracuda, performed once at startup rather than per frame. The reference avatar generated from the input video and the 8-count beat inserted into the audio complete the learning content. The affordance is what carries the learning argument: it turns an abstract video demonstration into spatially grounded cues on the learner's own body.","core_discovery":"The central claim is that dance learning can be personalized end-to-end: a learner picks a video of a desired dance, and the system converts it into structured AR content rather than displaying the video alone. The conversion pipeline estimates 3D body pose from the 2D footage, renders a reference avatar, embeds a beat counted in 8-count phrases into the extracted audio, and generates affordances by separating the avatar mesh into hand and foot regions. The affordances are then normalized once to the learner's body size using real-time pose estimation from the webcam and displayed over the learner's mirror image beside the reference avatar. The authors hold that these visual cues communicate movement mechanics and range of motion directly, addressing the known difficulty of learning upper-lower body coordination and fast beat synchronization from traditional videos.","pith_inferences":["If the one-time body-size normalization holds during fast, large-range dance motions, the same affordance pipeline could be applied to other motor-learning domains such as sports drills or rehabilitation exercises where joint trajectories matter.","A natural extension the paper leaves implicit is making the affordances reactive: instead of a static overlay, cues could change color or intensity when the learner's tracked wrists and ankles deviate from the reference motion, turning the system from passive content into live feedback.","The untested step is the accuracy of the normalization transform; a quantitative comparison of one-time versus per-frame alignment during dance movements would show whether the simplicity assumption costs accuracy.","If the visual affordance claim is right, an isolated affordance-on versus affordance-off comparison in a user study should show measurable differences in joint-angle error, not just self-reported preference."],"forward_implications":["Users can learn a chosen choreography without waiting for instructor-made content, since any accessible dance video is converted automatically.","Learners keep their eyes on the display and their own mirrored body at once, removing the device-switching interruption that self-learners reported.","Hand and foot affordances target the joints that prior work found hardest to learn, giving a principled reason to expect faster mastery of those movements.","Playback speed, repeat, and section navigation let learners slow difficult 8-count phrases, matching the pacing of a dance class without a teacher.","Because guidance happens on a large display rather than a headset, the approach avoids the fatigue and disorientation associated with prolonged HMD use."],"supporting_citations":[{"why":"supplies the 3D human pose and mesh from which the reference avatar and affordances are built.","marker":"[18]"},{"why":"provides the real-time 24-keypoint pose estimation used to normalize affordances to the learner's body.","marker":"[10]"},{"why":"gives prior evidence that real-time feedback on wrists and ankles aids dance motor learning, motivating the affordance placement.","marker":"[11]"},{"why":"converts the WHAM avatar mesh to FBX so vertex groups can be separated into affordance meshes in Blender.","marker":"[14]"},{"why":"establishes the 8-count segmentation used to structure the learning content and beat track.","marker":"[9]"}],"fun_headline_variants":["Turn any dance video into AR-guided steps","AR dance coach overlays moves from any video","Personalized AR dance lessons from any clip","Visual cues turn any dance video into a lesson","AR system syncs hand/foot cues to any dance clip"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system's benefit rests on the untested assumption that the one-time body-size normalization aligns the visual cues with the learner's actual moving body accurately enough to guide, rather than mislead, their movements.","fun_headline_variants_meta":{"raw":{"variants":["Turn any dance video into AR-guided steps","AR dance coach overlays moves from any video","Personalized AR dance lessons from any clip","Visual cues turn any dance video into a lesson","AR system syncs hand/foot cues to any dance clip"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1265,"prompt_tokens":753,"completion_tokens":512,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":369,"completion_tokens_details":{"reasoning_tokens":440}},"tokens_in":369,"tokens_out":512,"duration_ms":4258,"temperature":1.0,"reasoning_tokens":440,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:32:19.264447+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the learner's webcam view during a practice session and compare the projected positions of the affordance overlays against the tracked wrist and ankle positions frame by frame; if the average alignment error grows with movement speed or dance range, the one-time normalization is not sufficient. A controlled comparison of learners practicing an 8-count phrase with affordances versus with a plain mirrored video, measuring joint-angle error against the reference pose, would settle whether the cues improve learning.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the real-time 24-keypoint pose estimation used to normalize affordances to the learner's body."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"gives prior evidence that real-time feedback on wrists and ankles aids dance motor learning, motivating the affordance placement."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"converts the WHAM avatar mesh to FBX so vertex groups can be separated into affordance meshes in Blender."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"establishes the 8-count segmentation used to structure the learning content and beat track."}],"review_version":1}