{"id":"f1c799ea-ee2a-485a-b3b5-5717a9bd1542","arxiv_id":"2412.03055","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An edge-cloud UAV inspection system with a new lightweight detector (EdgeAnt) and keyframe selection reduces end-to-end latency by 88.9% and reports 42.1% mAP on a custom antenna dataset.","lead":"The paper builds a drone-based AI system that detects illegal or faulty antennas by running a lightweight detector on the drone and uploading only selected keyframes to the cloud. It reports lower latency than cloud-only processing and a new lightweight detector, EdgeAnt, that fits on small edge computers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central accuracy claim rests on a misdefined metric: Eq. 23 counts false positives as 'FN,' measuring precision on uploaded frames while ignoring missed targets, and the 28.2% figure does not match Tables VI/VII.","rationale":"The central claim has two parts: latency reduction and accuracy increase. The 88.9% E2EL reduction is reproducible from Table VI (1012+7.9=1019.9 ms vs 62+51.3=113.3 ms) and is credible. EdgeAnt is partially independently checkable: the detector code is public, COCO numbers are plausible for a 3M-parameter detector, and FLOPs/FPS are internally consistent. The soft spot is the accuracy half of the same claim. The metric in Eq. 23 is not accuracy: it divides true positives by true positives plus uploaded false positives, so it is precision restricted to uploaded keyframes. KSA's mu=6 confirmation requirement means a genuine antenna that appears in fewer than six consecutive edge detections is never uploaded and is therefore invisible to the numerator and denominator. In an interference-inspection task, missing a real source is the highest-cost error, so reporting precision-as-accuracy overstates system benefit. The internal tables compound the problem: Table VI's relative gains are 4.6% (same detector) or 20.5% (vs the weakest CO baseline), not 28.2%; the only 1.28 ratio in the paper is 42.1 mAP / 32.9 CO 480p accuracy, which mixes incompatible metrics. This is an internal inconsistency, not merely a disagreement with community consensus. A concrete rerun with standard precision/recall/F1 and false-negative accounting would settle it. I therefore keep the reader's CONDITIONAL verdict: the architecture and latency results are promising, but the headline accuracy improvement must be recomputed and restated before the system claims can be accepted.","tokens_in":25900,"tokens_out":5642,"duration_ms":50881,"concrete_test":"Re-run the Section IV.D system evaluation on the same 40-second video with proper object-level metrics: recall = TP/(TP+FN) where FN counts ground-truth antennas whose tracks never satisfy the mu=6 criterion, precision = TP/(TP+FP), and F1, for CO, ECC, and ECC+ modes with identical detection thresholds. Also reproduce each accuracy entry in Tables VI and VII from the raw upload logs and compute relative changes row by row; if ECC+'s accuracy gain over CO is not 28.2% under any consistent baseline, or if recall drops materially as precision rises, correct the abstract and conclusion accordingly.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing weakness is the system-level accuracy metric, not just dataset representativeness. In Section IV.A.4 Eq. (23) defines Accuracy = TP/(TP+FN), but the text defines FN as 'the number of noninterference source target frames uploaded' — i.e., false positives. The quantity computed is precision on uploaded frames, not accuracy and not recall. Under KSA (Algorithm 1), only detections surviving pixel threshold tau=120 and track threshold mu=6 are uploaded; true antennas that are detected in fewer than mu consecutive frames are never uploaded and never enter the denominator. Consequently the reported 90.4% accuracy for ECC+ can rise while missed interference sources are silently discarded, which is exactly the failure mode an inspection system must avoid. The 'increases accuracy by 28.2%' claim is also internally inconsistent: Table VI shows 90.4 vs 86.4 (+4.6%) for the same detector, and 90.4 vs 75.0 (+20.5%) versus the CO YOLOv8-n+BotSort baseline; no row in Tables VI or VII yields +28.2%. The closest arithmetic match (42.1/32.9 ≈ 1.28) mixes detector mAP with the 480p CO-mode accuracy from Table VII, which would be a category error. Thus the main quantitative system-level benefit beyond latency is currently unsupported by the paper's own equations and tables.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an AIoT system for UAV-based detection of unauthorized antenna interference sources, using an edge-cloud collaboration mode (ECC+) in which all detection and tracking run on the drone's edge device and a keyframe selection algorithm (KSA) uploads only selected inference results. The core detection component is EdgeAnt, a lightweight detector built from the LHGNet backbone, the HetBiFPN neck, and a small-object enhancement layer, together with an IMU-aided tracker called AntSort. The detector is ablated on a custom antenna dataset, benchmarked on COCO, and deployed on Jetson Xavier NX and Raspberry Pi 4B. System-level experiments compare E2EL, accuracy, power, and bandwidth across CO, ECC, and ECC+ modes, and a multi-UAV path-planning simulation is included.","tokens_in":26224,"tokens_out":6783,"duration_ms":63736,"significance":"The detector part of the paper is the strongest contribution: EdgeAnt is carefully ablated component-by-component, and the COCO benchmark provides an independent check that the architecture generalizes beyond the custom antenna dataset. The deployment measurements on two edge devices and the public detector code are also valuable, and the reported real-time inference speeds are credible. However, the headline system-level accuracy claim is not currently supported by the paper's own equations and tables: Eq. (23) computes precision on uploaded frames rather than accuracy or recall, and the claimed 28.2% improvement does not match any row in Tables VI or VII. The KSA thresholds are also selected on the evaluation scenario without a held-out validation sequence. If the system-level metrics are corrected and validated on a separate flight, the paper would make a useful engineering contribution.","major_comments":[{"comment":"Equation (23) defines Accuracy = TP/(TP+FN), but the text immediately defines FN as \"the number of noninterference source target frames uploaded,\" which is false positives, not false negatives in the detection sense. The quantity computed is therefore precision on uploaded frames, not accuracy and not recall. True negatives are not counted at all, and true antennas that are detected in fewer than mu consecutive frames are never uploaded, so they never enter the denominator. Under this metric, ECC+ can report a higher \"accuracy\" simply by filtering out low-confidence uploads while silently discarding missed interference sources, which is exactly the failure mode an inspection system must avoid. Please replace this with a standard detection metric, such as recall or F1 over all ground-truth target frames in the test video, and recompute all system accuracy numbers.","section":"Section IV.D, Eq. (23)"},{"comment":"The abstract claims that compared with CO mode, ECC+ \"increases accuracy by 28.2%.\" This number is not supported by Tables VI or VII. For EdgeAnt+AntSort, Table VI gives 90.4 for ECC+ versus 86.4 for the CO mode with the same detector, a gain of 4.0 percentage points, and versus 75.0 for the CO YOLOv8-n+BotSort baseline, a gain of 15.4 percentage points. Table VII gives 90.4 versus 68.2 for 720p CO (+22.2 percentage points) and versus 32.9 for 480p CO (+57.5 percentage points). The only arithmetic that yields roughly 28% is 42.1/32.9, which would mix EdgeAnt's detector mAP with the 480p CO-mode accuracy, a category error. Please either identify the exact baseline and calculation behind the 28.2% claim or remove it.","section":"Abstract and Section IV.D, Tables VI and VII"},{"comment":"The KSA parameters tau=120 px and mu=6 frames are tuned on the evaluation scenario: tau is derived from the observed Yagi antenna size in that imagery, and mu is selected from the interval that appears effective in test videos of the same scenario. No held-out flight sequence is used to validate transfer, and no sensitivity analysis is reported. Because the reported system accuracy and latency gains depend on these thresholds, the gains may be overfitted to one particular altitude, camera, and antenna distribution. Please validate the thresholds on a separate flight with different settings and report system accuracy and E2EL as functions of tau and mu.","section":"Section IV.C, Fig. 14, and Section IV.D"},{"comment":"All detector accuracy and mAP figures are reported as single point estimates. With only 600 training and 200 validation images in the custom antenna dataset, differences between top-ranked models, such as 0.423 versus 0.419 mAP in Table II, may be within run-to-run noise. Please report means and standard deviations over multiple training runs or bootstrap confidence intervals, and state the number of runs used. This is needed before the \"state-of-the-art\" claim on the custom dataset can be assessed.","section":"Section IV.B, Tables II-V"}],"minor_comments":[{"comment":"The multi-UAV objective function contains weighting coefficients alpha_1 through alpha_5, but the paper never states their values or how they were chosen. The simulation results in Fig. 17 are therefore not reproducible. Please provide the coefficient values and a brief sensitivity analysis.","section":"Section III.B, Eqs. (2), (10), (12)"},{"comment":"The detector code is publicly released, but the custom antenna dataset is not. Since the central detector comparison is on this dataset and the system validation uses a single 40-second video, the paper would be substantially stronger if the dataset and test video are released, or if a detailed public benchmark is used as an additional check.","section":"Section IV.A.1"},{"comment":"There are several cross-reference errors: the COCO results are said to be \"shown in Table IV\" but appear in Table V, and the text \"Fig. 5 shows the final complete neck architecture\" refers to a figure that depicts the multi-UAV setup rather than the neck. Please correct these references.","section":"Section IV.C and Tables IV-V"}],"recommendation":"major_revision","confidential_remarks":"The detector contribution is credible and independently checked on COCO, and the deployment results are useful. The main barrier is the system-level accuracy claim: Eq. (23) is mis-specified and the 28.2% figure in the abstract appears to be a category error. These can be fixed within the manuscript's scope by redefining the metric and recomputing the system numbers, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The detector part of this paper is worth a look. EdgeAnt is a credible lightweight detector: the ablations are thorough, the COCO benchmark provides an independent check, and the Jetson/Raspberry Pi deployment numbers are useful. The code is public. That side of the work is real and reproducible.\n\nThe system-level evaluation is where things fall apart. Eq. (23) defines Accuracy = TP/(TP+FN), but the text says FN is “the number of noninterference source target frames uploaded”—that is a false positive count, not a missed detection. The metric is actually precision on uploaded frames. Under KSA, true antennas that do not survive the tracking threshold are never uploaded and never counted, so the reported 90.4% can rise while real targets are silently dropped. That is exactly the failure mode an inspection system must avoid.\n\nSecond, the 28.2% claim does not match the paper’s own tables. Table VI shows 90.4% versus 86.4% for the same detector, and versus 75.0% for the CO baseline—about +4.6% and +20.5%, not +28.2%. The closest arithmetic is 42.1/32.9, which would mix detector mAP with a different resolution’s accuracy, a category error.\n\nThird, the custom dataset is small and unreleased (600 training, 200 validation images), and the KSA thresholds are tuned on the evaluation video. There are no error bars anywhere.\n\nThe detector is a competent, incremental contribution. The system claims need a corrected metric, a derived or withdrawn 28.2% figure, and ideally a released dataset. As is, the central quantitative benefit beyond latency is unsupported.\n\nI would send this to review—the detector work deserves a serious look—but the reviewers should push hard on the system evaluation. If the authors fix the metric, reconstruct the accuracy comparison, and release the data, this could be a useful paper. Right now, the abstract oversells what the evidence supports.","headline":"The detector work is solid and worth reading, but the headline 28.2% accuracy claim rests on a misdefined metric and an internal inconsistency.","tokens_in":26764,"tokens_out":2108,"would_cite":false,"duration_ms":19789,"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":"A lightweight edge detector plus selective keyframe uploads lets drones find antenna interference sources in real time, cutting end-to-end latency by 88.9% versus cloud-only processing.","keywords":["antenna interference detection","UAV inspection","edge-cloud collaboration","keyframe selection","lightweight object detection","tracking-by-detection","real-time inference","AIoT"],"falsifier":"Collect fresh aerial footage of the same three antenna types under different altitudes, cities, and lighting; run the released EdgeAnt code with the fixed KSA thresholds; if mAP drops substantially below 42.1% or latency savings shrink, the reported state-of-the-art and the 88.9% reduction are overfit to the single evaluation scene.","tokens_in":25708,"feed_emoji":"📡","tokens_out":7201,"duration_ms":59252,"temperature":0.7,"pith_summary":"The paper claims that a small neural network running on an edge computer aboard a drone can locate unauthorized or malfunctioning antennas well enough for real inspection work, and that uploading only carefully selected keyframes to a cloud server instead of streaming all video cuts end-to-end latency by 88.9% while raising accuracy by 28.2% compared with cloud-only processing. The detector, EdgeAnt, uses about three million parameters and reaches 42.1% mean average precision on a custom antenna dataset, while running at 21.1 frames per second on a Jetson Xavier NX. The system adds a tracker, AntSort, that feeds drone inertial measurements into a Kalman filter, and a keyframe selection algorithm with thresholds $\\tau=120$ pixels and $\\mu=6$ frames that decides which detections deserve a cloud upload. If these numbers hold, radio monitoring agencies could inspect antennas from the air with much lower latency than today's cloud-streaming approach.","feed_headline":"Drone edge AI cuts antenna-inspection latency by 88.9%","feed_subtitle":"A 3M-parameter detector on a Jetson-class board matches bigger models and uploads only keyframes, not video.","key_machinery":"The system's load-bearing mechanism is the combination of the EdgeAnt detector, the AntSort tracker, and the keyframe selection algorithm (KSA). EdgeAnt is a single-stage detector whose backbone LHGNet prunes and reconfigures HGNetv2, whose neck HetBiFPN mixes dual-path and tri-path feature fusion with heterogeneous convolution kernels, and whose enhancement layer TSRBlock uses dilated convolutions to keep small antenna targets visible. AntSort takes the detector's boxes and injects IMU acceleration from the drone as a control input to a Kalman filter, replacing the usual RANSAC-based camera motion compensation. The KSA filters detections by a pixel-size threshold $\\tau=120$ and uploads only targets that have been tracked for more than $\\mu=6$ consecutive frames, so each confirmed interference source is reported to the cloud once rather than continuously.","core_discovery":"The central claim is that an optimized edge-cloud collaboration mode, in which all detection and tracking happens on the UAV and only selectively chosen keyframe results are uploaded via MQTT, gives real-time antenna interference source localization without sacrificing accuracy. The paper reports that EdgeAnt, a single-stage detector built from a lightweight backbone (LHGNet), a heterogeneous bidirectional feature pyramid neck (HetBiFPN), and a small-object enhancement layer (EL) with two-step residual blocks, achieves 42.1% mAP on the authors' antenna interference source dataset with 3.0 million parameters and 14.7 GFLOPs, and 38.9% mAP on COCO with 5.4 GFLOPs. Deployed with TensorRT on Jetson Xavier NX it sustains 21.1 FPS at $1088 \\times 1088$ input, and with NCNN on Raspberry Pi 4B it sustains 4.8 FPS at $640 \\times 640$. In the full system, the keyframe selection algorithm reduces end-to-end latency from 1012 ms in cloud-only mode to 62 ms of communication plus 51.3 ms of inference in ECC+ mode, an 88.9% reduction, while accuracy rises from 75.0% to 90.4%.","pith_inferences":["Because the KSA thresholds are fixed in pixels and frames, a natural testable extension is to make them adaptive to flight altitude and camera focal length; the paper's own data suggest target pixel size varies with viewing angle.","The reported 28.2% accuracy improvement over cloud-only mode is measured on uploaded frames; an independent evaluation should separate detector recall from upload policy to see whether the gain is real detection improvement or simply fewer false uploads.","The same edge-cloud collaboration pattern with selective keyframes could apply to other aerial small-object inspection tasks, such as utility insulator defects or rooftop solar panel faults, since the architecture is task-agnostic."],"forward_implications":["Detector performance transfers across hardware: on Jetson Xavier NX with TensorRT the EdgeAnt pipeline sustains over 20 FPS at $1088 \\times 1088$, the only tested detector to meet real-time inspection needs.","The ECC+ mode keeps end-to-end latency under 150 ms even for 1080p video, and the influence of uplink bandwidth fluctuations shrinks relative to cloud-only mode.","With the KSA, each confirmed interference source is uploaded once rather than streamed continuously, so communication cost and false-positive traffic drop while accuracy rises.","The multi-UAV path planning formulation indicates that increasing fleet size from one to five UAVs reduces per-UAV energy and communication delay until airspace congestion offsets the gain."],"supporting_citations":[{"why":"Supplies HGNetv2, the backbone that LHGNet prunes and reconfigures.","marker":"[39]"},{"why":"Provides the dual-path fusion idea that HetBiFPN adapts.","marker":"[40]"},{"why":"Provides the weighted bidirectional feature pyramid and the EfficientDet baseline used for comparison.","marker":"[41]"},{"why":"YOLOv10, the main baseline and ablation reference whose architecture EdgeAnt is compared and contrasted with.","marker":"[48]"},{"why":"DWRSeg, the source of the two-step residual dilation idea used in the TSRBlock enhancement layer.","marker":"[50]"},{"why":"BotSort, the baseline tracker that AntSort modifies by injecting IMU acceleration.","marker":"[51]"},{"why":"COCO, the public benchmark used to demonstrate EdgeAnt's generalizability.","marker":"[53]"},{"why":"EC2Detect, the prior edge-cloud detection system whose division of labor the ECC+ mode revises by moving all inference to the edge.","marker":"[25]"}],"fun_headline_variants":["Edge AI on drones: 88.9% lower latency for antenna checks","Keyframe selection makes drone antenna inspection real-time","Tiny 3M-param model brings real-time antenna detection to drones","UAV edge-cloud mode cuts latency 88.9% and lifts accuracy 28%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results depend on the custom antenna dataset (600 training and 200 validation images with about 3,200 targets) being representative of real interference-source imagery, and on the KSA thresholds ($\\tau=120$ pixels and $\\mu=6$ frames) transferring to other flights and cameras.","fun_headline_variants_meta":{"raw":{"variants":["Edge AI on drones: 88.9% lower latency for antenna checks","Keyframe selection makes drone antenna inspection real-time","Tiny 3M-param model brings real-time antenna detection to drones","UAV edge-cloud mode cuts latency 88.9% and lifts accuracy 28%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001186,"raw_usage":{"total_tokens":5028,"prompt_tokens":1210,"completion_tokens":3818,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":826,"completion_tokens_details":{"reasoning_tokens":3738}},"tokens_in":826,"tokens_out":3818,"duration_ms":25095,"temperature":1.0,"reasoning_tokens":3738,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:48:46.857688+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect fresh aerial footage of the same three antenna types under different altitudes, cities, and lighting; run the released EdgeAnt code with the fixed KSA thresholds; if mAP drops substantially below 42.1% or latency savings shrink, the reported state-of-the-art and the 88.9% reduction are overfit to the single evaluation scene.","supporting_citations":[{"cited_title":"Path aggregation network for instance segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the dual-path fusion idea that HetBiFPN adapts."},{"cited_title":"Efficientdet: Scalable and efficient object detection,","cited_arxiv_id":null,"evidence_quote":"Provides the weighted bidirectional feature pyramid and the EfficientDet baseline used for comparison."},{"cited_title":"Dwrseg: Rethinking efficient acquisition of multi-scale contextual information for real-time semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"DWRSeg, the source of the two-step residual dilation idea used in the TSRBlock enhancement layer."},{"cited_title":"Microsoft coco: Common objects in context,","cited_arxiv_id":null,"evidence_quote":"COCO, the public benchmark used to demonstrate EdgeAnt's generalizability."},{"cited_title":"Ec²detect: Real- time online video object detection in edge-cloud collaborative iot,","cited_arxiv_id":null,"evidence_quote":"EC2Detect, the prior edge-cloud detection system whose division of labor the ECC+ mode revises by moving all inference to the edge."}],"review_version":1}