{"id":"e7332218-48d5-4f59-bfc8-b266db4ddf73","arxiv_id":"2506.08974","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A new open dataset of 30,440 underwater frames of 892 diving gestures, with synchronized glove and acoustic logs, recorded in sea and pool by ten divers at 1, 2 and 3 meters.","lead":"The paper presents an open dataset of over 30,000 underwater images of divers performing hand gestures, recorded in pool and sea with a synchronized sensor glove. It also reports glove recognition and acoustic transmission statistics; the dataset is intended as a benchmark for visual diver-robot gesture recognition.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Snippet extraction around glove detection time likely includes the preceding 'look at me' protocol gesture, contaminating visual labels for the main gestures.","rationale":"The reader correctly identified the glove-detection timing as a weak assumption, but the more specific and load-bearing failure mode is the snippet-window contamination by the protocol's mandatory preceding 'look at me' gesture. This is not just a timing offset; it is a structural label error that affects the visual content of the snippets regardless of whether Eq. (1) is exactly correct. The paper's own Table 4 shows that every main gesture is bracketed by 'look at me' gestures, so a 2.25 s pre-detection window almost certainly captures part of the preceding gesture. The dataset's utility as a baseline for the eight main gestures (which account for only 204 of the 892 instances) is therefore questionable without frame-level verification. This concern is addressable through dataset inspection or re-annotation, so conditional acceptance is appropriate: the dataset may be released, but its labels must be verified or the snippet extraction protocol revised. I partially agree with the reader's weakest assumption because the fixed-delay concern is real but secondary; the primary weakness is the unverified visual content of the snippets.","tokens_in":11904,"tokens_out":4105,"duration_ms":42872,"concrete_test":"Download from the public dataset at least one snippet folder per main gesture (e.g., 'OK', 'Ear problem', 'Reserve') for several divers and distances. Visually inspect frames at t = -2.25 s, -1.5 s, and -0.75 s relative to the stored glove detection timestamp, and count how many of the first 25% of frames in each snippet clearly show the 'look at me' gesture (index and middle fingers extended) rather than the labeled gesture. If this occurs in more than 10% of inspected snippets, label contamination is systematic and the dataset requires re-annotation or tighter extraction windows. Independently, recompute Eq. (1) for a subset of trials using the raw acoustic arrival timestamps and the reported modem protocol timings (30 ms chirp + 75 ms header + 25 ms data + 200 ms parity) to verify that the 0.330 s constant matches the actual modem behavior.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The dataset's central claim is that it provides a reliable visual benchmark for diving-gesture recognition. That claim depends on each snippet folder containing frames of the labeled gesture and only that gesture. Section 3.3 defines the snippet as all frames from 2.25 s before to 0.75 s after the glove's detection time. However, the glove protocol (Table 4) requires the diver to perform a 'look at me' gesture immediately before the main semantic gesture. Because the 'look at me' gesture (index and middle fingers extended) is visually distinct from most main gestures, the first portion of every main-gesture snippet will likely show the preceding 'look at me' rather than the labeled gesture. The paper reports no frame-level manual verification; Section 3.3 says only that reviewing footage showed 3 s covers the integral movements 'in most of the cases.' Additionally, Eq. (1) reconstructs the glove detection time from acoustic arrival using a fixed 0.330 s protocol delay plus a 3/1520 s travel-time term. Any variation in glove on-board processing, modem startup, or actual sound speed shifts the center of every snippet systematically. If either issue holds, a classifier trained on these snippets learns contaminated or misaligned labels, undermining the claimed baseline value.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a new underwater diving-gesture dataset intended for training and benchmarking visual gesture recognition systems for diver–robot communication. The data were collected with a stationary underwater camera while divers performed CADDIAN-based gestures at distances of 1, 2, and 3 meters in both pool and sea environments. In parallel, the divers wore a smart glove that detects gestures with on-board sensors and transmits the recognized command acoustically. The dataset contains 892 gesture instances and over 30,000 frames, organized into 3-second snippet folders that were extracted around the glove-reported gesture detection time. The paper also reports glove recognition statistics such as reaction time, gesture execution time, transmission time, and recognition success rates. The authors position the dataset as a baseline for comparing visual and glove-based underwater gesture recognition.","tokens_in":12140,"tokens_out":4174,"duration_ms":42349,"significance":"If the dataset is reliable, it fills a genuine gap: publicly available underwater gesture datasets with synchronized camera and sensor data are scarce, and the CADDIAN gesture language provides a principled vocabulary. The detailed experimental setup, the inclusion of five divers per environment, and the publication of raw video and sensor logs are valuable contributions. The paper also reports quantitative glove performance statistics, which could be useful for system-level comparisons. However, the dataset's utility hinges on the quality of its visual labels, and the current methodological description leaves important questions about label correctness and circularity insufficiently addressed. With appropriate documentation and validation, this dataset could become a useful benchmark; in its present form, the central claim of a reliable visual baseline is not fully established.","major_comments":[{"comment":"The snippet extraction uses the glove's detection time as the sole ground-truth anchor, and no independent frame-level annotation is reported. Since the glove is the authors' own system (references [3] and [19]) and the dataset is explicitly positioned to compare visual methods against the glove, this creates a circularity that undermines the claim that the dataset provides an independent visual benchmark. The paper should either provide a human-annotated validation subset or clearly demonstrate that the glove-derived timestamps are accurate enough to serve as ground truth, for example by reporting agreement with manual gesture boundaries on a sample of videos.","section":"Section 3.3"},{"comment":"The snippet window starts 2.25 seconds before the glove detection time. Given the glove protocol in Table 4, the diver must perform a 'look at me' gesture immediately before the main semantic gesture. Consequently, a large portion of each snippet likely contains the preceding 'look at me' rather than the labeled gesture. The paper only states that reviewing footage showed 3 seconds covers the movements 'in most of the cases,' without quantifying how often or to what extent the preceding gesture appears in the window. This contamination directly affects the label fidelity for the main gestures. The authors should quantify the fraction of snippets that include the preceding gesture and either adjust the window or annotate the active gesture interval within each snippet.","section":"Section 3.3 and Table 4"},{"comment":"The gesture detection time T_g is reconstructed from the acoustic arrival time T_a using a fixed protocol delay of 0.330 seconds and a fixed travel time of 3/1520 seconds. This assumes that the on-board processing time, modem startup, and sound speed are constant across trials. Any variation in these quantities will systematically shift the center of every snippet and thus misalign frame labels. The paper reports no uncertainty analysis for this timing reconstruction. The authors should quantify the expected timing error (e.g., by comparing multiple trials or using the redundant acoustic receivers) and discuss its impact on the snippet alignment.","section":"Equation (1), Section 3.2"},{"comment":"The dataset contains only 205 instances of the eight main semantic gestures, with per-class counts ranging from 17 to 37. In contrast, the auxiliary gestures 'Look at me' and 'Start communication' account for 688 instances. If the dataset is intended as a training or evaluation corpus for visual recognition of the main gestures, these numbers are quite small and highly imbalanced. The paper should either provide additional instances, clearly report the small per-class sizes as a limitation, and/or supply a recommended train/test split and evaluation metric that account for the imbalance, so that the baseline claim is scoped appropriately.","section":"Section 3.5, Table 3"},{"comment":"The glove recognition statistics in Table 5 and Figures 11–12 (recognition success rates, reaction times, transmission times) are reported without any error bars, confidence intervals, or per-user breakdown for the success rates. Since the paper presents these as benchmarks, the variability across the ten divers and across pool/sea conditions should be quantified. At minimum, the authors should report standard deviations or interquartile ranges for the timing metrics and the number of trials underlying the success rates.","section":"Section 3.6"}],"minor_comments":[{"comment":"The column header 'Totoal gesture time' contains a typo; it should read 'Total gesture time.'","section":"Table 5"},{"comment":"The 'Turn' gesture description contains '180°degrees' with a duplicated degree symbol; please correct to '180 degrees.'","section":"Table 1"},{"comment":"The paragraph states the video is stored in .mkv format 'along with frame timestamped data,' but it is not clear whether the frame timestamps are embedded in the video or stored in a separate file. This should be clarified for users who need to synchronize with the .csv logs.","section":"Section 2.4"},{"comment":"The synchronization procedure for the video is described only as recording 'a microsecond timer ... at multiple time difference samplings of each mission and averaged.' This is too vague to be reproducible. Please specify the exact method used to map video frame timestamps to the acquisition PC clock.","section":"Section 3.2"},{"comment":"The figure caption says 'Relative distribution of image samples per gesture by distance' but the y-axis and the exact definition of 'relative distribution' are not explained in the text. Please describe how the percentages were computed and whether they are based on frames or instances.","section":"Figure 7"},{"comment":"The sentence 'The first mentioned gesture starts the gesture recognition process on the smart diving glove and it is a gesture that cannot be easily done by accident during regular diving activities' is awkwardly phrased and should be revised for clarity. Additionally, the paper does not specify the finger numbering convention used in Tables 1 and 2, which would help readers interpret the descriptions.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript describes a potentially useful dataset, but the label-construction process is insufficiently validated and likely contaminated by the preceding 'look at me' gesture. The circular dependence on the glove as both label source and comparison baseline is a substantive concern for a dataset paper. I recommend asking the authors to add independent annotation or validation, quantify timing uncertainties, and provide baseline visual recognition results to demonstrate the dataset's utility. The dataset itself may be valuable even with these limitations, provided they are transparently documented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a genuinely useful dataset—single-camera underwater diving gestures, over 30,000 frames, nearly 900 snippets, with synchronized glove and acoustic logs, from five divers in both pool and sea at three distances. That combination is not in the CADDY stereo set, and the field could use more public data. The collection protocol is described in enough detail to reproduce, and the raw video plus metadata CSV give researchers a way to dig in themselves.\n\nThe soft spot is the snippet labeling. The glove protocol requires a 'look at me' gesture immediately before the main semantic gesture. The snippet is defined as all frames from 2.25 s before to 0.75 s after the glove detection time. Since the glove detection time corresponds to the main gesture, the first part of each main-gesture snippet will show the preceding 'look at me' rather than the labeled gesture. The paper says reviewing footage showed 3 seconds covers the integral movements 'in most of the cases,' but it does not report per-frame manual verification or any attempt to crop out the preparatory gesture. So the benchmark is noisier than the text suggests—a classifier trained on these snippets is learning a sequence that contains a different gesture, not a clean static pose.\n\nThere is also the circularity: the visual labels come from the glove detector itself, and missed gestures are excluded. That is a real weakness, but not fatal—the raw video is available and users can re-annotate if needed. The recognition statistics lack error bars, which is minor for a dataset paper. The timing reconstruction via Eq. (1) assumes a fixed protocol delay; any variation shifts the snippet center, adding more uncertainty.\n\nOverall, the dataset is worth having, but the labeling methodology needs a second pass before it can serve as a reliable baseline. The authors should either provide per-frame annotations or a tighter window that excludes the preceding gesture, or show that contamination is negligible with manual checks on a sample. As it stands, I would treat it as a starting point rather than a gold standard.\n\nRecommendation: send to peer review, but expect major revision on the labeling methodology. A careful referee should ask for a sample of snippets with frame-level labels to verify what is actually inside the folders.\n\nBest,\n[Your name]","headline":"Useful new underwater gesture dataset, but the snippet labeling likely includes the preceding 'look at me' gesture, so treat it as a starting point rather than a gold standard.","tokens_in":12651,"tokens_out":7531,"would_cite":false,"duration_ms":67236,"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":"The paper releases 30,440 labeled underwater gesture frames from 892 instances as a visual baseline for diver-robot communication.","keywords":["underwater hand gesture recognition","diver-robot communication","gesture recognition glove","CADDIAN gestures","acoustic communication","underwater imaging dataset","benchmark dataset","human-robot interaction"],"falsifier":"Measure the true delay by sending a test glove detection through the same acoustic chain while recording a synchronized electrical pulse; if the receiver timestamp minus the true detection time differs from $0.330 + 3/1520$ s by more than one frame at 30 fps, the published snippet boundaries are misaligned.","tokens_in":11736,"feed_emoji":"🤿","tokens_out":7742,"duration_ms":81292,"temperature":0.7,"pith_summary":"This paper presents an open dataset designed to give visual gesture recognition a fair baseline in underwater diver-robot communication, where acoustic glove commands are reliable but slow. Ten divers, five in the sea and five in a pool, performed eight CADDIAN diving gestures plus two protocol gestures at 1, 2, and 3 meters from a fixed camera, yielding 30,440 frames organized into 892 gesture-instance snippets. Each snippet is labeled by a smart glove that detects the gesture on board and timestamps it, so the visual data come with a sensor-derived ground truth rather than hand-drawn boxes. The glove's reaction times, gesture times, transmission delays, and detection rates are also reported, allowing vision algorithms to be compared directly with the glove-and-acoustics pipeline.","feed_headline":"A 30,440-frame underwater gesture dataset gives vision a baseline","feed_subtitle":"Smart-glove and acoustic logs let visual diver-robot communication be measured against sensor detection","key_machinery":"The object that carries the dataset is the gesture-detection glove plus the timing equation that turns its acoustic transmissions into video labels. The glove embeds dielectric-elastomer stretch sensors on each finger and an IMU on the back of the hand; when it recognizes a gesture, it confirms haptically and acoustically transmits a command. Equation (1) reconstructs the gesture time $T_g$ by subtracting a fixed $0.330$ s protocol delay and a $3/1520$ s acoustic travel time from the receiver timestamp $T_a$. That reconstructed time defines a 3-second snippet spanning from 2.25 s before to 0.75 s after detection, which becomes one labeled instance for machine learning.","core_discovery":"On its own terms, the paper's contribution is a public dataset and the claim that it can serve as a baseline for visual diving-gesture recognition. The dataset holds 30,440 frames organized into 892 gesture-instance snippets, recorded in the sea and in a pool by ten divers at 1, 2, and 3 m from a single wall-mounted camera. The ground truth is not manual annotation of the images: snippets are aligned to detection times from a smart glove, with the gesture time $T_g$ recovered from the acoustic arrival time $T_a$ via $T_a = T_g + 0.330\\,\\mathrm{s} + 3/1520\\,\\mathrm{s}$. The paper also reports the glove's recognition statistics, including an average detection rate near 85% and an acoustic reception rate of 79.83%, to serve as reference numbers for comparison with vision-based methods.","pith_inferences":["The paper does not itself train a recognizer; a natural next check is whether a simple visual classifier reproduces the glove's timing labels from the frames alone, which would validate the 2.25 s before and 0.75 s after window.","Because the metadata splits instances by environment, distance, and diver, one can test whether recognition accuracy drops more with distance or with turbidity, a comparison the paper reports the raw material for but does not perform.","If the fixed acoustic delay in Eq. (1) is wrong by a few tens of milliseconds, retiming the snippets with a corrected offset could shift benchmark results, so the timing assumption deserves direct measurement."],"forward_implications":["Visual recognition systems can be trained and evaluated on 30,440 raw 1920x1080 frames of ten gestures without needing the glove hardware, using the snippet folders as labels.","Because the same trials contain glove detection timestamps and acoustic reception timestamps, the dataset supports a head-to-head comparison of vision speed and accuracy against the reported glove detection rate near 85% and acoustic reception rate of 79.83%.","The 3-second snippet structure preserves temporal context, so dynamic gestures such as \"turn\" can be studied as frame sequences rather than only as static images.","The distance and environment metadata let future benchmarks quantify how visibility conditions affect recognition, since the images are deliberately not color-corrected or preprocessed."],"supporting_citations":[{"why":"Supplies the glove's gesture-detection algorithm and the wearable-sensing setup that produces the detection timestamps used to cut snippets.","marker":"[3]"},{"why":"Defines the CADDIAN gesture language and its semantics; the ten gestures recorded in the dataset are selected from this alphabet.","marker":"[16]"},{"why":"Introduces the earlier CADDY stereo dataset and the colored glove markers whose visual design this recording replicates.","marker":"[17]"},{"why":"Describes the glove-based hand gesture recognition with elastomer sensors and IMU that runs in parallel with the camera.","marker":"[19]"},{"why":"Defines the acoustic modem's protocol structure whose fixed delays set the 0.330 s term in the timing equation Eq. (1).","marker":"[24]"}],"fun_headline_variants":["Underwater gesture dataset: 30k frames, 10 divers","Diver-robot hand signals get a shared dataset","Dataset links visual and glove-based diver gestures","Baseline for underwater gesture recognition released"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The dataset's frame labels assume that every glove detection reaches the acoustic receiver after exactly $0.330$ s of protocol delay plus a $3/1520$ s acoustic travel time; if that delay varies between trials, the annotated snippets are systematically offset.","fun_headline_variants_meta":{"raw":{"variants":["Underwater gesture dataset: 30k frames, 10 divers","Diver-robot hand signals get a shared dataset","Dataset links visual and glove-based diver gestures","Baseline for underwater gesture recognition released"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1261,"prompt_tokens":992,"completion_tokens":269,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":208}},"tokens_in":608,"tokens_out":269,"duration_ms":3361,"temperature":1.0,"reasoning_tokens":208,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:57:46.509083+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the true delay by sending a test glove detection through the same acoustic chain while recording a synchronized electrical pulse; if the receiver timestamp minus the true detection time differs from $0.330 + 3/1520$ s by more than one frame at 30 fps, the published snippet boundaries are misaligned.","supporting_citations":[{"cited_title":"Ferreira, I","cited_arxiv_id":null,"evidence_quote":"Supplies the glove's gesture-detection algorithm and the wearable-sensing setup that produces the detection timestamps used to cut snippets."},{"cited_title":"Chiarella, M","cited_arxiv_id":null,"evidence_quote":"Defines the CADDIAN gesture language and its semantics; the ten gestures recorded in the dataset are selected from this alphabet."},{"cited_title":"Gomez Chavez, A","cited_arxiv_id":null,"evidence_quote":"Introduces the earlier CADDY stereo dataset and the colored glove markers whose visual design this recording replicates."},{"cited_title":"Sherlock, N","cited_arxiv_id":null,"evidence_quote":"Defines the acoustic modem's protocol structure whose fixed delays set the 0.330 s term in the timing equation Eq. (1)."}],"review_version":1}