{"id":"84b72ac2-686a-440a-bcfc-bb5f740a6766","arxiv_id":"2501.06651","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DeepLabV3+ outperforms Dynamic U-Net and PSPNet for aerial segmentation in a new 90-image Granada dataset, and parked cars can be identified by a fourth class or by a heuristic post-processing rule.","lead":"The authors train and compare three standard deep learning segmentation models on aerial images of Granada to locate parked cars, moving cars, and roads. They build a new 90-image labeled dataset for the city and report that DeepLabV3+ performs best, with parked car detection handled either by a separate model class or by a simple image-processing rule.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Parked-car detection claim rests solely on one qualitative example; no quantitative metric supports the 'favorable results' conclusion.","rationale":"The reader's weakest assumption focuses on dataset label quality and manual split representativeness. I agree these are relevant, but the most load-bearing gap is more direct: the paper's headline claim about parked-car detection is never quantitatively evaluated. Section 5.5 presents only one example image (Figure 4) and a general statement that both approaches yield favorable results. There is no error metric for the parked-car task, no comparison across all test images, and no validation of the heuristic's parameters. This is not a matter of label accuracy; even with perfect labels, the absence of measured performance means the claim is unsupported. The model comparison (DeepLabV3+ best) is at least supported by Table 1, though without error bars or multiple runs. Given this, the appropriate verdict remains CONDITIONAL: the authors should add quantitative parked-car evaluation and resolve the metric inconsistencies before the central claim can be accepted. The reader's verdict is therefore unchanged, but the condition should be sharpened to focus on the missing parked-car metrics rather than only dataset quality.","tokens_in":10326,"tokens_out":3415,"duration_ms":31227,"concrete_test":"Compute quantitative parked-car detection metrics on the 10 GranadaAerial test images for both pipelines: per-class IoU, precision, recall, and F1 for the 'parked car' class, using the provided ground truth. For Algorithm 1, vary the dilation kernel size (e.g., 5, 9, 15, 21, 31) and the background-vs-road threshold to assess sensitivity. If the best parked-car IoU is close to zero or not substantially better than a trivial majority-class baseline, the 'favorable results' claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim about parked-car detection—that both the heuristic approach and the dedicated four-class model yield 'favorable results'—is not supported by any quantitative evaluation in Section 5.5. Figure 4 shows a single test image with error masks, but no per-class IoU, precision/recall, or pixel accuracy is reported for the parked car class over the 10 test images. The heuristic (Algorithm 1) has unvalidated parameters (a fixed 15x15 dilation kernel and a 'background pixels count greater than road pixels count' threshold) with no sensitivity analysis, and its output is never compared against the four-class model numerically. Even if the GranadaAerial labels were perfectly accurate, this evidence gap leaves the parked-car claim as a qualitative, anecdotal assertion rather than a measured result. Additionally, the metric definitions in Section 5.4 are internally inconsistent: the Foreground Accuracy formula includes TN while claiming background is not taken into account, and the Dice-to-Jaccard conversion is applied without clarifying whether results are per-class averaged or computed globally. These issues directly weaken the strongest claim that parked cars can be reliably detected by either method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses parked-car detection in Granada from aerial imagery using semantic segmentation. The authors introduce a custom 90-image GranadaAerial dataset, train three models (Dynamic U-Net, PSPNet, DeepLabV3+) on UDD5 and UAVid, fine-tune the best model on Granada, and propose two parked-car detection approaches: a heuristic contour-dilation post-processing step applied to a car/road segmentation, and a direct four-class model that distinguishes parked from moving cars. The reported validation metrics in Table 1 indicate that DeepLabV3+ outperforms the other two models on validation loss, foreground accuracy, Dice, and Jaccard. The paper's stronger claim, that both parked-car detection approaches yield favorable results, is supported only by a single qualitative test image in Figure 4.","tokens_in":10576,"tokens_out":3847,"duration_ms":39803,"significance":"If the results hold, the paper would provide a useful comparison of standard segmentation architectures for parking-space detection in a new urban aerial dataset, and it would show that a simple contour-dilation heuristic can compete with a dedicated four-class model when labeled data are scarce. The authors deserve credit for making the code and the GranadaAerial dataset publicly available and for grounding the training pipeline in established public datasets, which improves reproducibility. However, the quantitative evidence is uneven: the segmentation-model comparison is supported by Table 1, but the main parked-car detection objective is never scored numerically, and the metric definitions in Section 5.4 contain inconsistencies. These gaps currently prevent the paper from substantiating its central claims.","major_comments":[{"comment":"The central claim that both parked-car detection approaches yield 'favorable results' rests on a single test image. No numerical evaluation over the 10 test images is reported for either the heuristic method or the four-class parked-car model: there are no per-class IoU, Dice, precision/recall, or error counts for the parked-car class, and no comparison between the two approaches. Algorithm 1's fixed 15x15 dilation kernel and the background-vs-road pixel threshold are validated only implicitly, with no sensitivity analysis. Because parked-car detection is the paper's stated primary objective, this is a load-bearing evidence gap; the authors should add quantitative test-set metrics, including per-class results for parked cars, and a sensitivity study for Algorithm 1's parameters.","section":"Section 5.5, Figure 4; Section 6"},{"comment":"The Foreground Accuracy formula ACC = (TP + TN) / (TP + TN + FP + FN) includes TN in both numerator and denominator, yet the prose states that 'background is not taken into account.' If background pixels are excluded, TN is either zero or undefined; if they are included, the metric is ordinary pixel accuracy rather than a foreground-focused metric. This contradiction makes Table 1 ambiguous. The Dice-to-Jaccard conversion JI = DSC / (2 - DSC) is also presented without clarifying whether the reported DSC and JI are macro-averaged over classes; the nonlinear conversion is only guaranteed to hold per class, not for macro averages. The authors should specify the exact computation (background excluded or not, macro vs micro averaging) and recompute or verify the reported values accordingly.","section":"Section 5.4, Metrics"},{"comment":"Table 1 reports validation metrics, but no test-set metrics are reported for the three segmentation models or for the fine-tuned Granada models. Since the dataset contains only 90 images with 10 validation and 10 test images, and the split was performed manually, it is unclear whether DeepLabV3+'s advantage in Table 1 transfers to held-out Granada images. The authors should report test-set metrics for all models and describe the label-annotation process in more detail, including any quality checks or inter-annotator agreement for the author-created GranadaAerial labels.","section":"Section 5.1 and Table 1"}],"minor_comments":[{"comment":"The color-coded error explanation is self-contradictory: the text first says green signifies false positives and red false negatives, but then states 'Green indicates areas present in the ground truth mask but not predicted, while red indicates areas predicted but not present,' which reverses the two definitions. This should be corrected to match the standard false-positive/false-negative convention.","section":"Figure 4 caption and Section 5.5 text"},{"comment":"The Dice formula's variable listing says 'TP, FP and FP are the same as for foreground accuracy,' where the second 'FP' should be 'FN.' This appears to be a typo, but it adds confusion to an already ambiguous metric definition.","section":"Section 5.4, Dice formula"},{"comment":"The manuscript contains several typographical and grammatical errors, including 'reffer', 'aditional', 'approppiate', 'achive', 'hte', and 'UA Vid' instead of 'UAVid.' A careful proofreading pass is needed.","section":"Throughout"},{"comment":"The description of the one-cycle policy and learning-rate finder is somewhat repetitive across the two fine-tuning approaches; consolidating this text and stating the actual learning-rate values used would improve clarity and reproducibility.","section":"Section 5.3"},{"comment":"The GranadaAerial dataset is distributed via a Google Drive link in footnote 2; the authors should consider a persistent repository or versioned release to ensure long-term availability, and they should specify the annotation schema and class definitions more precisely.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is modest in scope and appears to be a well-structured student project. The main technical concern is not the methodology but the evaluation: the paper's title and abstract promise parking-space detection, yet the quantitative evidence stops at segmentation metrics and one qualitative example. I believe this is fixable within the manuscript's scope by adding test-set evaluations and correcting the metric definitions. If the authors can provide those, the paper could be suitable for publication. The dataset and code availability are positive factors, though the dataset's small size and single-city focus should be acknowledged more explicitly as limitations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing worth knowing about this paper is that its actual contribution is the GranadaAerial dataset—90 manually labeled aerial images of Granada with car/road/background labels, plus parked-car labels in the test set—and a cleanly described transfer-learning pipeline that compares Dynamic U-Net, PSPNet, and DeepLabV3+. That part is credible. The models are standard, the training protocol (pretrain on UDD5/UAVid, fine-tune on Granada) is sensible, and Table 1 gives consistent evidence that DeepLabV3+ wins on validation loss, foreground accuracy, Dice, and Jaccard. The authors also ship code and links to the data, which makes the work reproducible. I believe them.\n\nThe soft spot is exactly what the stress-test flags: the parked-car detection result, which is the stated objective, is never scored. Section 5.5 shows one test image and asserts that both the heuristic and the four-class model \"accurately represent\" the parked cars, and the conclusion repeats \"favorable results.\" No per-class IoU, precision/recall, or even a count over the ten test images. The heuristic in Algorithm 1 has a fixed 15x15 kernel and a background-vs-road pixel threshold with no sensitivity analysis. That is a small dataset, but with only 10 test images, a per-image or per-class number would have been cheap to produce. Its absence makes the central claim anecdotal.\n\nThere are also two internal inconsistencies worth fixing. The Foreground Accuracy formula includes TN while the prose says background is not taken into account; that formula is just accuracy. And the Dice-to-Jaccard conversion is stated without saying whether values are per-class averaged or global. Minor, but they matter for a benchmark paper.\n\nI do not see a load-bearing flaw in the segmentation comparison itself. The manual split is a reasonable choice for 90 images. The domain-shift discussion is honest. The paper is limited and the evaluation gap is real, but the work is not misleading about the architecture novelty—it claims a dataset and a comparison, and it delivers those. The \"favorable results\" line overreaches.\n\nWho it's for: someone working on aerial parking detection or small-city transfer learning might find the dataset and pipeline useful. It deserves a serious referee, mainly because the dataset is new and the code is public; a referee can ask for the missing numbers. I'd send it to review, but with a note that the parked-car evaluation must be quantitative before publication.","headline":"A modest but honest dataset-and-benchmark paper; the segmentation comparison is fine, but the headline parked-car claim is backed by one image and no numbers.","tokens_in":11056,"tokens_out":1907,"would_cite":false,"duration_ms":17761,"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":"DeepLabV3+ gives the strongest segmentation of parked cars, moving cars, and roads in aerial images of Granada, and parked cars can be found either by a dilation heuristic or by adding a fourth class.","keywords":["semantic segmentation","aerial imagery","parking space detection","DeepLabV3+","PSPNet","Dynamic U-Net","GranadaAerial dataset","transfer learning"],"falsifier":"Have an independent annotator re-label the 10 GranadaAerial test images, then re-run the final fine-tuning and evaluation on the re-labeled set; if DeepLabV3+ no longer leads on Dice or Jaccard, or if the two parked-car approaches change their relative quality, the paper's central claim is not stable. A field-level test on a new set of, say, 30 aerial images taken on different days and weather conditions would also settle whether the favorable parked-car results generalize.","tokens_in":10169,"feed_emoji":"🚗","tokens_out":8736,"duration_ms":81492,"temperature":0.7,"pith_summary":"This paper claims that DeepLabV3+ is the most promising of three tested architectures for segmenting parked cars, moving cars, and roads in aerial images of Granada, Spain. It builds and uses a custom 90-image dataset, GranadaAerial, labeled by the authors, and shows that pretraining on two existing aerial datasets followed by fine-tuning on Granada produces favorable results. The paper also demonstrates two routes to parked-car detection: a post-processing contour-dilation heuristic applied to a car/road segmentation, and a model trained to output a dedicated parked-car class. If correct, this offers a low-cost path to parking-occupancy monitoring from aerial imagery in cities with little existing labeled data.","feed_headline":"DeepLabV3+ leads aerial parking detection in Granada","feed_subtitle":"A 90-image dataset and a contour-dilation trick let aerial photos reveal parking occupancy.","key_machinery":"The central mechanism is DeepLabV3+'s encoder-decoder design with atrous spatial pyramid pooling (ASPP), which captures multi-scale context using dilated convolutions and is paired with a ResNet101 backbone. The empirical argument rides on a two-stage transfer schedule: pretraining all three models on UDD5 and UAVid, then fine-tuning the selected model on GranadaAerial. The parked-car distinction is carried by Algorithm 1, a contour-dilation heuristic whose 15x15 dilation kernel extends each detected car contour so that the pixel-type composition of its surroundings (background versus road) decides whether the car is parked.","core_discovery":"After pretraining Dynamic U-Net, PSPNet, and DeepLabV3+ on a fusion of the UDD5 and UAVid datasets, DeepLabV3+ with a ResNet101 backbone achieves the lowest validation loss (0.05404) and the highest foreground accuracy (0.7726), Dice coefficient (0.7955), and Jaccard index (0.6836) at the 50th epoch. The authors then fine-tune DeepLabV3+ on the custom GranadaAerial dataset, and report that both parked-car detection approaches yield favorable results: one applies Algorithm 1, a heuristic that dilates car contours with a 15x15 kernel and labels a car as parked when its dilated neighborhood contains more background pixels than road pixels; the other trains the network to output a fourth parked-car class directly.","pith_inferences":["An implicit testable extension is to replace the binary background-versus-road decision in Algorithm 1 with a richer neighborhood context (curbs, sidewalks, vegetation), since cars parked on wide road shoulders may be misclassified when road pixels dominate the dilated region.","Because the dataset is small, self-training on unlabeled Granada orthophotos using the fine-tuned model's high-confidence predictions as pseudo-labels could cheaply expand GranadaAerial and improve the parked-car class.","The paper compares the two parked-car approaches mainly on a single illustrative test image; reporting per-image error metrics over the full test set would let readers see exactly where the heuristic breaks and whether the four-class model is consistently better.","A further consequence, not pursued in the paper, is that the same three-class segmentation plus heuristic could be adapted to other Mediterranean cities with similar street geometry, while parking garages and underground lots would require new classes."],"forward_implications":["Parking occupancy in Granada could be monitored over time by running the trained model on periodically captured aerial images, without installing street-level sensors.","DeepLabV3+ becomes the natural baseline architecture for future aerial semantic-segmentation work on small custom urban datasets.","The contour-dilation heuristic offers a viable alternative to training a dedicated parked-car class when labeled examples of parked versus moving cars are scarce.","The UDD5 and UAVid pretraining transfers to a geographically different city after a short fine-tuning phase, suggesting the same recipe may work in other urban environments.","Adding a fourth parked-car class lets the network directly output parking locations, at the cost of slightly higher classification complexity than the three-class plus heuristic pipeline."],"supporting_citations":[{"why":"Defines the DeepLabV3+ encoder-decoder with atrous separable convolution and ASPP, the architecture the paper selects as strongest.","marker":"[6]"},{"why":"Defines PSPNet with pyramid pooling, one of the two competing baseline architectures.","marker":"[44]"},{"why":"Provides the aerial-image segmentation adaptation of U-Net that Dynamic U-Net builds on.","marker":"[19]"},{"why":"Introduces the U-Net encoder-decoder architecture from which Dynamic U-Net is derived.","marker":"[30]"},{"why":"Supplies the UDD5 aerial dataset used for the initial pretraining of all three models.","marker":"[7]"},{"why":"Supplies the UAVid drone-image dataset used together with UDD5 for pretraining.","marker":"[26]"},{"why":"Provides the aerial orthoimagery of Granada from which the custom GranadaAerial dataset is built.","marker":"[20]"},{"why":"Introduces dilated convolutions, the core mechanism behind DeepLabV3+'s multi-scale ASPP module.","marker":"[42]"}],"fun_headline_variants":["DeepLabV3+ best for aerial parking detection in Granada","Aerial AI spots parking in Granada with DeepLabV3+","Granada parking detection: DeepLabV3+ leads the pack","DeepLabV3+ tops Granada parking maps from drone imagery"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole comparison rests on the 90 hand-labeled GranadaAerial images being accurate and representative of Granada's parking scenes; if those labels are inconsistent or unrepresentative, the reported model ranking and parked-car results do not transfer to real use.","fun_headline_variants_meta":{"raw":{"variants":["DeepLabV3+ best for aerial parking detection in Granada","Aerial AI spots parking in Granada with DeepLabV3+","Granada parking detection: DeepLabV3+ leads the pack","DeepLabV3+ tops Granada parking maps from drone imagery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1410,"prompt_tokens":915,"completion_tokens":495,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":422}},"tokens_in":531,"tokens_out":495,"duration_ms":4928,"temperature":1.0,"reasoning_tokens":422,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:55:09.478674+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have an independent annotator re-label the 10 GranadaAerial test images, then re-run the final fine-tuning and evaluation on the re-labeled set; if DeepLabV3+ no longer leads on Dice or Jaccard, or if the two parked-car approaches change their relative quality, the paper's central claim is not stable. A field-level test on a new set of, say, 30 aerial images taken on different days and weather conditions would also settle whether the favorable parked-car results generalize.","supporting_citations":[{"cited_title":"Mapping urban green spaces at the metropolitan level us- ing very high resolution satellite imagery and deep learning techniques for semantic segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the aerial-image segmentation adaptation of U-Net that Dynamic U-Net builds on."},{"cited_title":"Large-scale structure from motion with semantic con- straints of aerial images","cited_arxiv_id":null,"evidence_quote":"Supplies the UDD5 aerial dataset used for the initial pretraining of all three models."},{"cited_title":"Uavid: A semantic segmentation dataset for uav imagery","cited_arxiv_id":null,"evidence_quote":"Supplies the UAVid drone-image dataset used together with UDD5 for pretraining."},{"cited_title":"PNOA Image of Granada, Spain, 2022","cited_arxiv_id":null,"evidence_quote":"Provides the aerial orthoimagery of Granada from which the custom GranadaAerial dataset is built."}],"review_version":1}