{"id":"86c43cc3-b22b-42fd-a3a5-03fc6c8ccbf5","arxiv_id":"2505.00995","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A 1.4 kg indoor UAV with LiDAR-inertial odometry and 3D multi-object tracking estimated cherry tomato count and weight in a GNSS-denied greenhouse, reaching 94.4% counting and 87.5% weight accuracy on a single 13.2 m lane.","lead":"A small drone with LiDAR, a depth camera, and an inertial sensor flew down a cherry tomato greenhouse aisle without GPS and used 3D tracking of individual tomatoes to estimate the harvest. On one 13-meter lane, it produced 94 tracked fruits where 89 were harvested, and it estimated average fruit weight to within about 12 percent, in a roughly 10-second flight.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (11) does not reproduce its own calibration points: evaluating the printed cubic at h = 35, 40, 42 mm yields 22.9/27.4/32.3 g, not 13.5/18.1/23 g, so the reported 87.5% weight accuracy is not reproducible from the manuscript.","rationale":"The reader's weakest-assumption identification points to Eq. (11) and the three-point fit, which is exactly the load-bearing element for the weight-estimation claim. My read agrees with this focus, but sharpens it: the printed coefficients do not even reproduce the three calibration points that allegedly motivated the fit. This is a stronger, more objective flaw than the underdetermination argument alone, because it indicates a numerical error or a mismatch between the reported equation and the experiment. The counting accuracy (94.4%) is grounded in a harvest-based validation and is less affected by this issue, so the overall verdict need not change from CONDITIONAL. The paper's contributions in hardware integration, LiDAR-inertial odometry, and 3D MOT remain plausible; the weight-estimation pipeline needs correction and re-validation. Recommending UNCHANGED reflects that the proposed revision path remains conditional on fixing the mapping and either supplying a justified model or more calibration data with error bars. The concrete test—evaluating Eq. (11) at the stated calibration heights—is a decisive, low-cost check that either confirms the typo or exposes a deeper methodological gap.","tokens_in":10476,"tokens_out":6586,"duration_ms":60245,"concrete_test":"Compute Eq. (11) at h = 35, 40, and 42 mm. Since the outputs (22.9, 27.4, 32.3 g) mismatch the stated calibration weights (13.5, 18.1, 23 g), request the corrected coefficients or raw track-height data from the authors. Then re-run the forward-view weight estimation: average the predicted weights from Eq. (11) over the estimated tracks and compare to the ground-truth 19.14 g average. The 87.5% accuracy claim is valid only if this recomputation yields an error near 12.5%.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section III-D, Eq. (11) is introduced as a cubic mapping from tracked tomato height to weight, fitted to three harvested fruits (35 mm/13.5 g, 40 mm/18.1 g, 42 mm/23 g). Evaluating the printed polynomial at these heights gives 22.9 g, 27.4 g, and 32.3 g respectively—errors of 9.4 g, 9.3 g, and 9.3 g. The equation does not pass through any of the stated calibration points, so it cannot be the mapping described. This internal inconsistency means the 87.5% weight-estimation accuracy in Table I cannot be reproduced from the manuscript: either the coefficients are misprinted, the calibration data are misreported, or the actual mapping used for the forward-view run was different. The reader's additional point that a cubic has four coefficients while only three points are given compounds the problem: even if the printed equation were a valid interpolation, the chosen coefficients would be one of infinitely many, and no rationale (physical density model, intercept constraint, or held-out data) is provided to justify it. Since the headline yield claim depends on this weight mapping, the weight component of the central claim is unsupported in its current form.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a lightweight 1.4 kg quadcopter for indoor, GNSS-denied greenhouse monitoring. The UAV fuses a Livox MID-360 LiDAR, an Intel RealSense D405 RGB-D camera, and an IMU, using GLIM LiDAR-inertial odometry for pose estimation. Cherry tomatoes are detected with YOLOv10, represented as 3D cubes from depth images, and associated across time by a 3D multi-object tracker. The authors report a 94.4% counting accuracy and an 87.5% weight-estimation accuracy on a 13.2 m harvesting lane flown in 10.5 s, validated by post-flight harvesting of 89 tomatoes weighing 1704.4 g. A second growing-row dataset is used for a qualitative tracking analysis of unripened tomatoes, where only a 42.5% positive-track rate is observed. The paper claims that the system demonstrates the potential of UAVs for efficient robotic yield estimation in commercial greenhouses.","tokens_in":10812,"tokens_out":3367,"duration_ms":36751,"significance":"If the reported numbers are reproducible, the paper makes a useful systems contribution: it integrates off-the-shelf perception and odometry components into a small UAV and validates yield estimates against real post-harvest ground truth on a commercial greenhouse lane. The authors are transparent about the manual control mode, the offline processing, and the exploratory nature of the unripened-tomato experiment. The platform details and the 10.5 s inspection time for a 13.2 m lane are concrete and practically relevant. However, the weight-estimation claim currently rests on an internally inconsistent calibration equation and an evaluation protocol that uses the same lane for fitting and validation, so the headline accuracy numbers should not be accepted without a corrected and properly validated analysis.","major_comments":[{"comment":"The printed cubic mapping does not reproduce the stated calibration points. Evaluating weight = 0.00178 h^3 + 0.00993 h^2 - 7.36 h + 192 at h = 35, 40, and 42 mm gives approximately 22.9 g, 27.4 g, and 32.3 g, respectively, rather than the reported 13.5 g, 18.1 g, and 23 g. Because the 87.5% weight accuracy in Table I is computed through this mapping, the central weight-estimation claim cannot currently be reproduced from the manuscript. Please correct the coefficients, the calibration data, or the description, and make the calibration data or script available so the reader can verify the mapping.","section":"Section III-D, Eq. (11)"},{"comment":"The weight mapping is fitted to three harvested tomatoes from the validation lane and then evaluated on that same lane, and a cubic polynomial has four coefficients while only three data points are provided. Unless an additional constraint (for example, a zero-intercept condition, a physical density model, or independent data) is specified, the coefficients in Eq. (11) are not uniquely determined. In addition, the association and filtering thresholds (dist_max, update weights w_p and w_v, minimum association count, volume cutoff, maximum accepted depth, and coordinate constraints) are all described as empirically tuned on the same dataset. This evaluation protocol cannot support a generalizable accuracy claim. Please provide a held-out test lane, or at minimum a cross-validation scheme, and report the sensitivity of the results to the tuned thresholds.","section":"Section III-D and Section IV-B"},{"comment":"The headline claim of 94.4% counting accuracy and 87.5% weight accuracy is based on a single forward-view run. The tilted-view run on the same lane produces 77.6% counting accuracy and a 36.1% weight error, which the authors attribute to higher flight velocity and limited depth range, but these factors are not quantitatively isolated. The conclusion and abstract should either restrict the accuracy claim to the forward-view configuration or support it with repeated trials across varied speeds and camera poses. As written, the abstract's phrasing 'achieves 94.4% counting accuracy and 87.5% weight estimation accuracy' overstates the evidence.","section":"Table I and Section IV-B"}],"minor_comments":[{"comment":"The phrase 'two dataset' should be 'two datasets'.","section":"Abstract and Introduction"},{"comment":"The manuscript uses inconsistent spacing in 'UA V' and 'UAV'; please choose one convention and apply it consistently.","section":"Throughout"},{"comment":"The caption says 'Pink allows shows the UAV's flight path'; this should likely read 'Pink arrows show the UAV's flight path'.","section":"Fig. 2 caption"},{"comment":"The table mixes the ground-truth row with flight configurations without clear separation; please clarify that the 'True' row is the post-harvest ground truth and define how the percentage errors are computed (for example, absolute relative error on count and on average weight).","section":"Table I"},{"comment":"The statement that the maximum accepted depth of 1 m 'exceeding the D405 specification' should be clarified, since exceeding a sensor's specified range typically means relying on noisier measurements rather than a hard failure threshold.","section":"Section IV-B"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle is not the system concept but the validity of the reported numbers: Eq. (11) is internally inconsistent and the validation uses the same lane for fitting and evaluation. These issues are fixable with corrected data and a proper train/validation split, so I am recommending major revision rather than rejection. The paper would also benefit from a clear statement of which experiments are confirmatory and which are exploratory, since the growing-row result is explicitly qualitative."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things before spending time on this one. First, the integrated UAV system actually flies in a GNSS-denied greenhouse and the counting result is grounded by a real post-flight harvest. Second, the headline weight accuracy is not reproducible from the manuscript: the cubic mapping in Eq. (11) does not fit the three calibration points the text says it was fitted to. Evaluating it at 35, 40, and 42 mm gives 22.9, 27.4, and 32.3 g, not 13.5, 18.1, and 23 g. That is an internal contradiction, and it guts the 87.5% weight claim.\n\nWhat is new: a compact 1.4 kg quad with Livox MID-360, RealSense D405, and Jetson Orin NX running GLIM for LiDAR-inertial odometry, adapted cube-based 3D MOT for cherry tomatoes, and a field trial on a commercial lane. The counting accuracy of 94.4% on a 13.2 m lane in 10.5 s is a meaningful data point, and the authors are honest about the tilted-view run being much worse (77.6% count, 36.1% weight error) and about validation being only partial.\n\nThe soft spots are significant. Eq. (11) is load-bearing and broken: three points cannot determine a four-coefficient cubic, and the printed coefficients don't come close to the stated data. On top of that, all association and filtering thresholds (dist_max, update weights, min associations, volume cutoff, max depth, coordinate constraints) are hand-tuned on the same dataset. There is no held-out test, no baselines, and no error bars. The growing-row tracking result is 42.5%, which the authors concede is poor. So the generalizable yield-estimation claim is thin.\n\nWho is this for? Robotics researchers working on indoor UAV monitoring or agricultural perception will find the platform design and the failure analysis useful, but they should not quote the weight accuracy. The paper deserves a serious referee because the system and counting result are real and the issues are fixable. My recommendation: major revision. Fix or substantiate Eq. (11), add a held-out evaluation (or at least cross-validation), and report the mapping's original data. If the mapping correction is straightforward, a revised version could be a solid applied paper.","headline":"The UAV platform and counting result are real, but the weight-estimation claim is not reproducible: Eq. (11) doesn't pass through its own calibration points.","tokens_in":11344,"tokens_out":3295,"would_cite":false,"duration_ms":31307,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a 1.4 kg quadrotor navigating by LiDAR-inertial odometry in a greenhouse without satellite signals can count ripened cherry tomatoes with 94.4% accuracy and estimate their weight with 87.5% accuracy during a…","keywords":["UAV yield estimation","greenhouse monitoring","GNSS-denied navigation","LiDAR-inertial odometry","3D multi-object tracking","cherry tomato","RGB-D perception"],"falsifier":"Collect the per-fruit height and weight for all 89 harvested tomatoes, fit Eq. (11) using only a randomly chosen subset (for example, three fruits), and compute the average weight error on the remaining fruits; if the held-out error exceeds the reported 12.5%, the weight-estimation claim does not generalize.","tokens_in":10285,"feed_emoji":"🍅","tokens_out":9793,"duration_ms":88503,"temperature":0.7,"pith_summary":"The paper is trying to establish that a small aerial robot can carry out pre-harvest yield estimation inside a commercial greenhouse where satellite positioning is unavailable, a job normally given to ground robots that need rails, lifts, and wide aisles. It reports a 1.4 kg quadrotor that flies down a 13.2-meter cherry-tomato row in about 10.5 seconds, using LiDAR-inertial odometry (laser-plus-inertial motion estimation) to hold its position and a 3D multi-object tracker to turn RGB-D images into tomato counts and sizes. In the harvesting-row experiment, the system counted 94 ripened fruits against a true count of 89 (94.4% counting accuracy) and estimated the average fruit weight within 12.5% of the harvested mean (87.5% weight accuracy). The authors also show that a tilted camera view and faster manual flight degrade both metrics, and that green, heavily occluded fruit is tracked at only 42.5% counting accuracy, which they use to motivate future viewpoint planning.","feed_headline":"94.4% of cherry tomatoes counted in a 10.5-second drone flight","feed_subtitle":"A 1.4 kg UAV, using laser-inertial positioning and 3D tracking, estimates tomato yield without a satellite signal.","key_machinery":"The load-bearing mechanism is a cube-based 3D multi-object tracking pipeline. A YOLOv10 detector gives 2D boxes; for each box, the median depth inside the projected region fixes $Z$, and the pinhole equations convert the box width and height into metric width $W$, height $H$, and length $L$, where $L = (W+H)/2$ exploits the near-sphericity of cherry tomatoes. GLIM, a LiDAR-inertial odometry estimator, supplies the 6D camera pose that transforms these cubes into a global coordinate frame. Tracks are matched to new detections when the Euclidean distance is at most 0.04 m; matched tracks update their cube state as a convex combination with weight 0.7; a track is accepted after three associations. Ripened tomatoes are those with volume above 12 cm³, and their height feeds Eq. (11), a cubic height-to-weight polynomial, to produce the yield estimate.","core_discovery":"The central claim is that one short forward-view flight of a lightweight quadrotor, with no GNSS and no farm infrastructure modifications, is enough data to estimate cherry-tomato yield. Each detected tomato is projected into the depth image to get a median distance, converted into a metric cube via the pinhole model, and placed in a global frame using the GLIM pose; cubes are then associated into tracks using a 0.04 m Euclidean threshold, updated as convex combinations, and emitted only after three associations. Ripened-fruit tracks with volume over 12 cm³ are counted, and track height is converted to weight by a cubic polynomial (Eq. 11). In the validating experiment, this pipeline produced 94 tracks versus 89 harvested fruits, and the average predicted tomato weight was 21.53 g against a true average of 19.14 g. The paper reports these numbers as 94.4% counting accuracy and 87.5% weight-estimation accuracy for a 13.2 m lane inspected in 10.5 seconds.","pith_inferences":["The cubic weight mapping in Eq. (11) is fitted to only three harvested tomatoes, yet a cubic has four coefficients, so those three points do not determine the curve; until the mapping is validated on independent fruit, the 87.5% weight accuracy should be treated as a proof-of-concept bound rather than a settled metric.","A volume- or density-based weight model calibrated on a larger sample across rows would likely reduce weight error and remove the need to re-fit per lane.","The single forward-view flight leaves open how double-counting scales with lane length; a full-row survey with per-fruit ground-truth matching would test whether the 94.4% count holds beyond one 13.2 m segment.","The 42.5% tracking accuracy on green fruit suggests that active viewpoint selection, planning where the UAV looks next, is the natural next lever, and a next-best-view planner along the flight path would be a concrete extension of this system."],"forward_implications":["Greenhouse yield monitoring can move off the ground: a single 13.2 m row is scanned in about 10.5 s without rails, lifts, or wheel modifications, which is the efficiency advantage claimed over UGV platforms.","The same RGB-D-plus-pose tracking pipeline can be adapted to other indoor crops with roughly spherical fruit, as long as the size-to-weight relation is re-calibrated for the new crop.","Because the yield computation is done offline from stored sensor streams, flight time is decoupled from compute time, so heavier tracking models could be applied without lengthening the flight.","The tilted-view and unripened-fruit experiments bound the method: counting accuracy falls to 77.6% with a tilted camera and to 42.5% for occluded green fruit, so viewpoint choice, not just detector quality, drives accuracy."],"supporting_citations":[{"why":"Supplies the LiDAR-inertial odometry estimator, GLIM, whose 10 Hz pose output allows the UAV to fly without GNSS and places tomato cubes in a global frame.","marker":"[21]"},{"why":"Provides the greenhouse 3D multi-object tracking approach the paper adapts, establishing the multi-view perception formulation for tomato fruits.","marker":"[22]"},{"why":"Cited with [22] as the multi-object tracking basis; its 3D single-shot detection and tracking with transformers is the comparison reference for building 3D fruit representations.","marker":"[23]"}],"fun_headline_variants":["Drone counts 94.4% of cherry tomatoes in 10.5s","UAV yield estimate in 10.5s: 94.4% count, 87.5% weight","No GPS? Indoor drone estimates tomato yield in 10.5s","GNSS-denied UAV counts cherry tomatoes at 94.4% accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported weight accuracy depends on a cubic height-to-weight curve fitted to just three tomatoes from the same lane, and three measured points cannot uniquely determine a cubic, so if that curve does not represent the other 86 fruits the 87.5% accuracy is an artifact.","fun_headline_variants_meta":{"raw":{"variants":["Drone counts 94.4% of cherry tomatoes in 10.5s","UAV yield estimate in 10.5s: 94.4% count, 87.5% weight","No GPS? Indoor drone estimates tomato yield in 10.5s","GNSS-denied UAV counts cherry tomatoes at 94.4% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001097,"raw_usage":{"total_tokens":4604,"prompt_tokens":994,"completion_tokens":3610,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":3517}},"tokens_in":610,"tokens_out":3610,"duration_ms":22580,"temperature":1.0,"reasoning_tokens":3517,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:29:16.265791+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect the per-fruit height and weight for all 89 harvested tomatoes, fit Eq. (11) using only a randomly chosen subset (for example, three fruits), and compute the average weight error on the remaining fruits; if the held-out error exceeds the reported 12.5%, the weight-estimation claim does not generalize.","supporting_citations":[{"cited_title":"GLIM: 3D range- inertial localization and mapping with GPU-accelerated scan matching factors,","cited_arxiv_id":null,"evidence_quote":"Supplies the LiDAR-inertial odometry estimator, GLIM, whose 10 Hz pose output allows the UAV to fly without GNSS and places tomato cubes in a global frame."},{"cited_title":"Development and evaluation of automated localisation and reconstruction of all fruits on tomato plants in a greenhouse based on multi-view perception and 3D multi-object tracking,","cited_arxiv_id":null,"evidence_quote":"Provides the greenhouse 3D multi-object tracking approach the paper adapts, establishing the multi-view perception formulation for tomato fruits."},{"cited_title":"MOT-DETR: 3d single shot detection and tracking with transformers to build 3d representations for agro-food robots,","cited_arxiv_id":null,"evidence_quote":"Cited with [22] as the multi-object tracking basis; its 3D single-shot detection and tracking with transformers is the comparison reference for building 3D fruit representations."}],"review_version":1}