{"id":"58753d4b-83d0-41dc-b2a6-95dc81848213","arxiv_id":"2506.21635","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A YOLO-style detection-plus-segmentation network warns of UAV landing deviations with 98.6% accuracy on a new 9,142-image dataset.","lead":"This paper presents a lightweight vision model that watches drone landing video and warns when the drone is drifting off its landing pad, reporting 98.6% warning accuracy and a 0.7 second warning delay. It also introduces a new landing-deviation dataset and an evaluation metric for warning speed.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"AWD metric excludes inference latency, so the reported 0.7 s warning delay is not a wall-clock delay; the headline real-time responsiveness claim is unsupported as stated.","rationale":"The reader's weakest assumption concerns the center-of-image label rule. That is a genuine external-validity threat, but it is tied to an explicit operational assumption (landing starts with the target centered), which the dataset may or may not satisfy. The AWD latency exclusion is a direct, internal flaw in the metric used for the strongest quantitative claim: the paper itself defines AWD as the time until the system issues a warning, then removes the interval during which the system is computing. No external assumption is needed to see that this makes 0.7 s not a wall-clock delay. The issue also affects the baseline comparison in Table IV, because slower models lose more latency. I agree with the reader's overall CONDITIONAL verdict but flag a different load-bearing soft spot; hence partial agreement. The fix is straightforward: report AWD including processing time, or clearly relabel the metric as a frame-level detection lag and avoid claiming real-time delay. Until that is done, the headline responsiveness figure is not interpretable as stated.","tokens_in":14637,"tokens_out":11867,"duration_ms":118511,"concrete_test":"Recompute AWD on the released UAVLandData videos (or request the authors to do so) with wall-clock warning times: for each frame i, let t_i be the capture timestamp, p_i be the model's inference time for that frame, and let P_i = t_i + p_i (or the completion time of the frame that first triggers the rule). Recompute AWD = mean|T_i - P_i| for AeroLite-MDNet and all baselines. If the corrected AWD for AeroLite-MDNet exceeds 1.0 s, or if the ranking among models changes, the reported 0.7 s AWD is not a valid real-time warning delay and the headline claim must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-D defines AWD as the average time between the actual onset of deviation and the warning, but then states that frames captured during the model's processing time are excluded and the last frame captured before inference completes is used instead. This assigns the warning to the capture timestamp of the input frame, omitting the inference latency k. A real-time warning is only issued after inference completes; the reported 0.7 s AWD therefore measures a frame-index perception lag, not the wall-clock warning delay. The omission is especially consequential on Jetson TX2 (7.3 FPS, k ≈ 0.137 s) and for slow baselines (DeepLabV3+ at 0.2 FPS, k ≈ 5 s), where it can understate AWD by seconds and change the comparative ranking. The central claim 'effective real-time ... warning' rests on this number, so it is not supported as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes AeroLite-MDNet, a lightweight multi-task network combining a detection head and a segmentation head for detecting deviations during UAV landing, together with a new dataset (UAVLandData) of 9,142 images annotated with bounding boxes and segmentation masks, and a new evaluation metric (Average Warning Delay, AWD). The system center-crops each frame to 640x640, runs detection and segmentation, and uses Algorithm 1 to issue a warning when the target's offset from the image center exceeds a threshold. The paper reports 98.6% warning accuracy, 0.7 s AWD, 7.12M parameters, 104.2 FPS on an RTX 3070, and 7.3 FPS on a Jetson TX2, and compares against six baselines in detection, segmentation, and warning-system metrics.","tokens_in":1819,"tokens_out":2455,"duration_ms":79499,"significance":"If the results hold, the contribution is practically useful: the model is compact, the multi-task detection-segmentation design is reasonable, the UAVLandData dataset is a potentially valuable community resource, and the paper explicitly commits to releasing code. The novelty is incremental rather than foundational, but the application is timely. However, the headline real-time claim rests on an AWD definition that explicitly excludes inference latency, and the ground-truth deviation onset and the decision threshold are not defined; as written, the experimental support for 'effective real-time warning' is incomplete and needs substantial revision.","major_comments":[{"comment":"The AWD metric as defined does not measure wall-clock warning delay. The text states that 'frames captured during the model's processing time are excluded. Instead, the last frame captured before inference completes is used.' This means the warning time Pi in Eq. (15) is the capture timestamp of an input frame, not the time at which the warning is issued after inference. For a system that warns only after inference finishes, the true delay is at least Pi + k, where k is the per-frame processing time. On the Jetson TX2, k is about 0.137 s at 7.3 FPS, and for DeepLabV3+ at 0.2 FPS, k is about 5 s, so the omission can understate AWD by seconds and change the relative ranking of methods. The abstract's 'AWD of 0.7 seconds' and the conclusion's claim that the system 'meets the real-time warning requirements' are therefore not supported as stated. The authors should report AWD including inference latency, or explicitly redefine the metric as a frame-index perception lag and avoid claiming real-time wall-clock responsiveness based on it.","section":"Section III-D, Eq. (15); Section IV-C, Table IV; Conclusion"},{"comment":"The ground-truth deviation onset time Ti in Eq. (15) is never defined. To compute AWD, the authors must specify how the 'actual deviation time' is determined for each landing sequence: is it manually annotated, is it derived from the true nest position crossing a threshold, and what is the protocol for the early phase when only houses are visible? Relatedly, the threshold delta in Algorithm 1 is never reported, yet ACC, FPR, and AWD all depend on it. Without this information, the central warning-system metrics are not reproducible. The paper should state the onset annotation protocol, the value of delta, and how the five landing stages in Fig. 6 are segmented in time.","section":"Section III-D, Eq. (15); Section III-C, Algorithm 1; Section IV-C"},{"comment":"All performance numbers are reported as single-point estimates with no error bars, no number of repeated runs, and no statement of the number of landing sequences n used in Eq. (15). Given that the test split contains only 1,636 images and the warning-system metrics are aggregate over entire landing trajectories, the authors should report variance across runs or folds, at least for the headline metrics (AWD, ACC, FPR). They should also state how many landing videos/sequences contribute to Table IV.","section":"Section IV-C, Tables III and IV"},{"comment":"The deviation rule in Algorithm 1 and the center-crop preprocessing both assume that the camera optical center is the true landing position and that the landing zone is centered in the field of view at the start of descent. If the camera is mounted with a known offset or the target is not centered initially, the labels and the AWD measure target-to-image-center alignment rather than actual landing deviation. The authors should validate this assumption (e.g., with camera calibration or by annotating the true touchdown point) and state what happens when the center-crop moves the target outside the field of view. A concrete test would be to compare warnings based on image-center distance with warnings based on a calibrated ground-truth landing point.","section":"Section IV-A and Fig. 2"}],"minor_comments":[{"comment":"The code availability URL in the abstract (https://github.com/ITTTTTI/Maskyolo.git) differs from the one in the first-page footnote (https://github.com/AeroLite-MDNet/AeroLite-MDNet.git); the two should be aligned.","section":"Abstract and Section I"},{"comment":"The CIoU formula appears malformed: the first term is written as rho^2(c2,c1) in both numerator and denominator, and the roles of alpha and v are not those of the standard CIoU loss. Please check the equation against the cited source and correct it.","section":"Section III-B1, Eq. (10)"},{"comment":"Equation (14) contains a dangling phrase 'where y_i is the true label' after an incomplete expression; the sentence 'where y_hat_i is the output, , and y_i is the true label' should be cleaned up.","section":"Section III-B2, Eq. (14)"},{"comment":"Algorithm 2's loop condition 'while accuracy of verification set > 90%' is inconsistent with the text that says 'We stop training until the model achieves an accuracy greater than 90%'; the condition should be inverted or clarified.","section":"Section III-B, Algorithm 2"},{"comment":"The focal-loss description states that 'When a sample is difficult to classify, p_t approaches 1', but for a difficult sample p_t is small, while for an easy sample p_t approaches 1 and the factor (1-p_t)^gamma reduces the loss weight. The sentence should be corrected.","section":"Section III-B1"},{"comment":"The header and body of Table III are misaligned in the typeset text; for example, the UNet row renders the FPS TX2 and FPS GPU values as '0.5243.9'. Please reformat the table so each column is unambiguous.","section":"Table III"},{"comment":"The sentence 'our false alarm rate is slightly higher than Mask R-CNN and lower than ConvNeXt-V2' contradicts the table, which shows FPR of 0.7 for both Mask R-CNN and the proposed method and 0.4 for ConvNeXt-V2; the text should be corrected.","section":"Section IV-C, Table IV"}],"recommendation":"major_revision","confidential_remarks":"The central problem is the definition of AWD: as written, the headline 0.7 s figure excludes the very latency that determines whether the system can warn in real time. This is fixable if the authors re-measure the metric with inference latency included, define T_i and delta, and add error bars. The paper's scope fits TCSVT as an application-oriented systems paper, but the timing claim and the reproducibility of the warning-system evaluation should be resolved before acceptance. Also note the duplicate GitHub URL mismatch, which should be handled before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi,\n\nThe paper's genuine contributions are the UAVLandData dataset and the AWD metric, but the reported 0.7s average warning delay excludes inference latency, so the real-time responsiveness claim is unsupported as written. The stress-test note is right: Section III-D explicitly drops frames captured during processing and uses the last capture timestamp instead of the warning-emission time. That makes the metric a perception lag, not a wall-clock system delay. For the Jetson TX2 at 7.3 FPS the missing k≈0.14s is modest; for DeepLabV3+ at 0.2 FPS it is ~5s and would reorder the baselines. So the headline number does not mean what the abstract says.\n\nWhat's actually new: the dataset with 9,142 real images across 24 scenarios is a real asset for drone-nest landing research, and the AWD idea is a good way to compare warning systems if it were defined and measured end-to-end. The model itself is a compressed YOLO variant with SCA/CA attention plus a mask head; each component is known, but the integration and the evaluation against segmentation and multi-task baselines is a legitimate engineering extension. The ablation is careful and shows the segmentation branch helps.\n\nSoft spots, in proportion:\n- AWD is the main problem. T_i (actual deviation onset) is undefined, threshold delta is never reported, and no error bars are given. The metric also uses absolute difference without saying whether early warnings are treated as errors. Fixing the latency definition is mandatory.\n- The code is not available yet; the abstract links to one repo, the full text to another. The dataset link is a placeholder. That blocks reproducibility.\n- The deviation rule and anchor boxes are derived from the same labeled data, so the evaluation is not independent from the labeling assumptions; the center-crop assumption (camera aligned to touchdown point) is a real limitation for generalization.\n- Baselines are a bit thin: no recent YOLO-family detectors with comparable size, and no Mask-RCNN variant with the same backbone. Not fatal, but the efficiency comparison would be more convincing.\n\nOverall, the paper deserves a serious referee pass; it is a plausible systems paper with a new dataset. I would send it to review but tell the authors to release the artifacts and re-report AWD as an end-to-end wall-clock delay, or rename the metric. My own verdict is conditional; with the metric corrected and the dataset accessible, this could be a useful reference for drone-nest landing systems. I would not cite it in the next 12 months until the artifacts are out.","headline":"The paper's new dataset and AWD metric are useful, but the reported 0.7s warning delay excludes inference latency, so the real-time claim is not supported as stated.","tokens_in":15334,"tokens_out":2878,"would_cite":false,"duration_ms":30520,"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":"AeroLite-MDNet claims a lightweight detection-plus-segmentation network can warn of UAV landing deviations in real time, with 98.6% warning accuracy and a 0.7-second average delay.","keywords":["UAV landing","landing deviation warning","multi-task learning","object detection","semantic segmentation","lightweight network","edge deployment","Average Warning Delay"],"falsifier":"Mount the camera with a fixed yaw offset so the landing pad is consistently, say, 40 pixels right of the image center at descent start, and run Algorithm 1 without recalibration; if the system still reports 98.6% warning accuracy and 0.7 s AWD on ground truth labeled by the true pad position, the center-alignment assumption is validated, and if not, the reported metrics are an artifact of that assumption.","tokens_in":14476,"feed_emoji":"🛬","tokens_out":7353,"duration_ms":67502,"temperature":0.7,"pith_summary":"Landing a UAV on a docking station often fails when GPS is weak or jammed, so pilots need a vision-based way to tell, in real time, that the aircraft is drifting off course. This paper proposes AeroLite-MDNet, a lightweight network that runs detection and segmentation together: detection finds the landing nest, QR code, or house, and segmentation recovers its shape, so the system can compare the target's position with the image center and decide whether to warn. To support this, the paper contributes UAVLandData, a 9,142-image landing dataset, and a new metric, Average Warning Delay (AWD), which measures how many seconds pass between the true start of a deviation and the system's warning. The reported result is 98.6% warning accuracy, a 0.7-second AWD, 7.12M parameters, 104.2 FPS on an RTX 3070, and 7.3 FPS on a Jetson TX2, which the authors present as evidence that real-time edge deployment is feasible.","feed_headline":"UAV landing network flags drift in 0.7 seconds","feed_subtitle":"A 7.12M-parameter model with detection and segmentation branches reaches 98.6% warning accuracy at 104 FPS.","key_machinery":"The carrying mechanism is a shared backbone-and-neck feature pyramid feeding two heads. The backbone uses CA Conv and SCA C3 attention modules to keep features compact while emphasizing relevant channels and spatial locations; the neck produces large, medium, and small feature maps. The detection head uses preset anchors with focal loss and CIoU loss, and the segmentation head fuses multi-scale features and combines Dice, BCEWithLogits, and focal losses, with the final mask cropped by the detected box to remove stray predictions. The decision rule (Algorithm 1) and the AWD metric complete the mechanism: AWD is the mean absolute difference between true deviation onset and warning time, with frames captured during model inference excluded so reported delay is not an artifact of processing speed.","core_discovery":"The central claim is that object detection and semantic segmentation can be combined in one small network to give a practical UAV landing-deviation warning. The detection branch supplies location and category for the landing target, and the segmentation branch supplies shape, which lets the system handle rotated targets that vertical bounding boxes cannot capture. Deviation is decided by Algorithm 1: if only a house is visible, the UAV deviates when the house is not at the image center; once the nest or QR code appears, the UAV deviates when the distance between the image center and target center divided by the nest's length exceeds a threshold. On the authors' data and baselines, the model reaches 98.6% warning accuracy and 0.7 seconds AWD while staying light enough (7.12M parameters, 104.2 FPS on GPU, 7.3 FPS on TX2) for near-real-time warning on embedded hardware.","pith_inferences":["The system's reference point is the image center, so field deployments need camera alignment with the true touchdown point; adding a calibration or adaptive reference would make the numbers transferable to other aircraft.","AWD is a generalizable metric for any time-critical visual warning task where ground-truth onset times exist, not only UAV landings.","The warning output is a monitor, not a controller; wiring this signal into the flight control loop is the natural next step, a direction the paper itself flags.","The ablation suggests segmentation is the largest single contributor to detection quality (recall from 80.6% to 94.1%), implying that the segmentation branch does real work rather than acting as a regularizer; this could be tested on other small-object aerial detection datasets."],"forward_implications":["A warning system of this size can run on embedded hardware (7.3 FPS on Jetson TX2), so landing monitoring need not depend on a ground-station GPU.","AWD gives the community a direct, interpretable measure of how fast a deviation warning arrives, not just whether the classification is correct.","The joint detection-segmentation design lets the system handle rotated and small landing targets better than either task alone, which matters in real descents.","The public dataset can serve as a benchmark for future vision-based landing-deviation research."],"supporting_citations":[{"why":"Mask R-CNN is the joint detection-segmentation baseline it must beat in accuracy and efficiency.","marker":"[28]"},{"why":"RTMDet provides the real-time detection-segmentation baseline and the CSPDarkNet-style design point the model is compared against.","marker":"[22]"},{"why":"DINO is the transformer detection baseline for the detection comparison.","marker":"[5]"},{"why":"The YOLO-style anchor-based detection head and loss setup the architecture builds on.","marker":"[16]"},{"why":"Focal loss supplies the classification and object loss in the detection head.","marker":"[23]"},{"why":"CIoU loss supplies the bounding-box regression loss in the detection head.","marker":"[24]"},{"why":"Dice loss is the main component of the segmentation loss.","marker":"[25]"},{"why":"UNet is a segmentation baseline whose low edge-device speed motivates the lightweight design.","marker":"[26]"},{"why":"DeepLabV3+ is a segmentation baseline compared on accuracy and parameter count.","marker":"[27]"}],"fun_headline_variants":["Lightweight UAV landing net flags drift in 0.7s","98.6% accurate landing deviation alert for UAVs","Multi-task model detects UAV landing drift with 98.6% accuracy","UAV landing alert: 0.7s delay, 98.6% accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything hinges on the assumption that the center of the camera image is the true landing position and that the landing zone starts inside the center crop; if the camera is misaligned with the actual touchdown point, the labels and the reported warning delay measure the wrong quantity.","fun_headline_variants_meta":{"raw":{"variants":["Lightweight UAV landing net flags drift in 0.7s","98.6% accurate landing deviation alert for UAVs","Multi-task model detects UAV landing drift with 98.6% accuracy","UAV landing alert: 0.7s delay, 98.6% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000597,"raw_usage":{"total_tokens":2794,"prompt_tokens":946,"completion_tokens":1848,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":1769}},"tokens_in":562,"tokens_out":1848,"duration_ms":13267,"temperature":1.0,"reasoning_tokens":1769,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:48:34.476211+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Mount the camera with a fixed yaw offset so the landing pad is consistently, say, 40 pixels right of the image center at descent start, and run Algorithm 1 without recalibration; if the system still reports 98.6% warning accuracy and 0.7 s AWD on ground truth labeled by the true pad position, the center-alignment assumption is validated, and if not, the reported metrics are an artifact of that assumption.","supporting_citations":[{"cited_title":"Mask r-cnn,","cited_arxiv_id":null,"evidence_quote":"Mask R-CNN is the joint detection-segmentation baseline it must beat in accuracy and efficiency."},{"cited_title":"ultralytics/yolov5: v6. 0-yolov5n’nano’models, roboflow integration, tensorflow export, opencv dnn support,","cited_arxiv_id":null,"evidence_quote":"The YOLO-style anchor-based detection head and loss setup the architecture builds on."},{"cited_title":"Enhancing geometric factors in model learning and inference for object detection and instance segmentation,","cited_arxiv_id":null,"evidence_quote":"CIoU loss supplies the bounding-box regression loss in the detection head."},{"cited_title":"V-net: Fully convolutional neural networks for volumetric medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Dice loss is the main component of the segmentation loss."},{"cited_title":"U-net: Convolutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"UNet is a segmentation baseline whose low edge-device speed motivates the lightweight design."}],"review_version":1}