{"id":"4ea34a47-ecca-4d61-ab5b-a415786102e6","arxiv_id":"2501.05076","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A ResNeXt-101 plus FPN model achieves very high fingertip segmentation accuracy on a private dataset, but the state-of-the-art claim is weakened by cross-dataset comparison and missing artifacts.","lead":"TipSegNet is a deep learning model that segments all fingertips from a single hand image, reporting 98.7% mean IoU and 99.9% accuracy on a private test set. The model combines a ResNeXt-101 backbone with a feature pyramid network, but the strong performance claim relies on comparisons to other methods evaluated on different datasets and lacks public code or data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline mIoU is computed with micro-averaging over a background-dominated 9-class task, so it is not the standard mean per-class mIoU and cannot support the claimed SOTA comparison.","rationale":"The reader flagged the private test set and split as the weakest point. I agree those are risks, but the most load-bearing issue is internal: the paper's own definition of mIoU is non-standard and background-dominated. This affects the headline claim even before any concern about data leakage or representativeness. The fix is concrete and modest: report class-averaged per-class IoU and compare baselines on the same protocol. Because this can be addressed, the appropriate verdict remains conditional rather than acceptance or rejection.","tokens_in":11821,"tokens_out":9817,"duration_ms":91869,"concrete_test":"Ask the authors to release the per-class confusion matrix on the 224-image test set and compute the macro-averaged mIoU (mean of the 9 per-class IoU values) and the mean of the 8 fingertip-class IoUs. If the fingertip-class mean is materially below 0.987, or if the macro mIoU drops below the Table III baseline mIoU values when those baselines are re-run on the same test set, the headline SOTA claim is not supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's quantitative SOTA claim rests on 'mIoU 0.987' and 'accuracy 0.999'. Section II-B states that 'we employ micro-averaging for all metrics', aggregating TPs/FPs/FNs across all classes before computing the metric. In the 9-class task (8 fingertip classes plus background), micro-averaged IoU is the global pixel IoU, not the standard mean-per-class mIoU used in the publications cited in Table III. Because background pixels dominate the image, micro-averaged IoU and accuracy are both inflated by the background class; a model with much lower fingertip-class IoU can still score near 0.99 on these micro-metrics. The comparison in Table III is also cross-dataset, with baseline numbers taken from other papers. Therefore the central claim that TipSegNet 'outperforms existing methods' is not established by the reported numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"TipSegNet proposes a ResNeXt-101 encoder with a Feature Pyramid Network decoder for segmenting eight fingertip classes plus background directly from whole-hand grayscale images. The authors train on a mix of 220 images from their prior work [29] and 2,037 in-house smartphone images, apply a stochastic augmentation pipeline, and report micro-averaged mIoU=0.987, accuracy=0.999, and F1=0.994 on a 224-image test set. Ablation studies varying augmentation strength and backbone (ResNet-34/50/101 vs. ResNeXt-101) are reported. The paper claims state-of-the-art performance over both whole-hand and single-finger segmentation methods.","tokens_in":11968,"tokens_out":6149,"duration_ms":60986,"significance":"If the reported numbers were computed with the same protocol as the cited baselines and on a fair test set, a deep model that segments all four fingertips in one pass with near-perfect overlap would be a practically useful contribution to contactless fingerprint preprocessing. The paper is honest about model-size trade-offs and includes ablation experiments, which is good practice. However, the significance is currently undercut by three issues: the headline metric is a micro-averaged global pixel IoU, not the standard mean-per-class mIoU used in the cited literature; the comparison table takes baseline numbers from other papers and other datasets; and the test-set construction is insufficiently described. No code or data is released, so independent verification of the central claim is not possible from the manuscript alone.","major_comments":[{"comment":"The abstract and conclusion claim state-of-the-art results based on mIoU=0.987 and accuracy=0.999, but Section II-B states that micro-averaging is used for all metrics. With nine classes of which background is likely the majority, micro-averaged IoU is the global pixel IoU, not the mean-per-class mIoU reported in the cited baseline papers. Consequently, the headline numbers are inflated by background agreement and are not comparable with the mIoU values in Table III. Please report macro-averaged per-class mIoU, per-class IoU for each of the eight fingertip classes, and accuracy computed on non-background pixels or with class-balanced aggregation, and explain how Eq. (1) is extended to the multi-class setting.","section":"Section II-B, Table III, Abstract"},{"comment":"None of the baseline methods in Table III were run on the authors' test set; the numbers are taken from the respective publications, which used different datasets, sensors, annotation protocols, and class definitions (e.g., Otsu and color-histogram methods segment only hand versus background, while U-Net/EfficientNet/SqueezeNet operate on single fingers). The bold 'highest scores' in each group therefore do not establish that TipSegNet 'outperforms existing methods.' To support the SOTA claim, the authors should either run the baselines on their own test set with the same evaluation code or restrict the claims to a descriptive comparison and explicitly discuss the non-comparability.","section":"Table III"},{"comment":"The data description is internally inconsistent: the text says 220 images from [29] plus 2,037 in-house images, totaling 2,257, but the stated split sums to 1,788 + 224 + 224 = 2,236, leaving 21 images unaccounted for. In addition, the paper does not specify whether the split is by subject or by image; if multiple images per subject exist in the 2,257-image pool, random image-level splitting can cause identity leakage and inflate the test metrics. The test set is only 224 images, and no confidence intervals or significance tests are provided. Please clarify the split procedure, account for all images, and report error bars or per-class variability.","section":"Section II-B"},{"comment":"The ablation studies are reported as single runs with no variance information, and all configurations achieve nearly identical micro-averaged metrics (Accuracy 0.997-0.999, IoU 0.976-0.987). The text nevertheless concludes that ResNeXt-101 'does offer improvements' and that augmentation 'plays a critical role in preventing overfitting,' but these conclusions are not supported by the table: the no-augmentation model achieves the same accuracy and IoU as the fully augmented model, and the discussion even notes that the no-augmentation training loss was lowest. Multiple seeds with mean and standard deviation, or a statistical test, are needed before attributing differences to the architectural or augmentation choices.","section":"Tables IV and V"}],"minor_comments":[{"comment":"The total parameter count is inconsistent with the sum of its parts: the encoder (8.264e8), decoder (2.608e6), and segmentation head (1,161) sum to about 8.29e8, but the table reports '828.965 * 10^8', which is off by orders of magnitude and does not match the 829-million figure used in the discussion.","section":"Table I"},{"comment":"The F1 score is reported as 0.993 in the Results section but as 0.994 in the Discussion and in Tables IV and V; please standardize the value.","section":"Section III and Discussion"},{"comment":"The text says all metrics are calculated on the test set, but Figure 4 shows results from the validation set; please clarify which set was used for the qualitative examples and for the ablation tables.","section":"Section II-B and Figure 4"},{"comment":"The training description gives one exact epoch count (853) and later refers to 'around 850 epochs' for the ablation backbones; reporting exact epoch counts and random seeds for all runs would improve reproducibility.","section":"Section II-A3"},{"comment":"The caption says 'class 0 describes the separation of the fingers from the background,' which is confusing because class 0 is the background class; rephrase to state that class 0 is the background label.","section":"Figure 4 caption"},{"comment":"The tables label the metric 'IoU' while the text calls it 'mIoU'; given that micro-averaging is used, the metric should be named consistently and its exact definition stated in every table caption.","section":"Section III, Table IV/V"}],"recommendation":"major_revision","confidential_remarks":"The core contribution is a reasonable engineering application of existing segmentation components, but the central performance claim is not supportable as written because of the non-standard metric and the non-comparable baseline table. The issues are fixable within the manuscript's scope: recompute standard per-class metrics, provide a subject-disjoint split and dataset accounting, and either run baselines on the test set or weaken the SOTA claim. If the authors choose not to provide these, the paper would be better positioned as a descriptive case study rather than a state-of-the-art comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: TipSegNet is a clean application of ResNeXt-101 + FPN to multi-fingertip segmentation from whole-hand images. The direct four-fingertip segmentation without a separate detection step is genuinely convenient, and the ablations are honest about saturation. But the headline “outperforms SOTA” claim is not supported by the numbers as reported.\n\nWhat's new and good: the architecture is standard (ResNeXt + FPN is a well-trodden combination), but applying it to this specific multi-class fingertip task on a whole-hand input is apparently new. The paper is clearly written, the training details are transparent (loss curves, epochs, compute), and the discussion of model size versus gain is sensible. The augmentation ablation shows no real benefit, and they say so rather than overselling it.\n\nSoft spots, in order of severity. First, the stress-test concern is correct: Section II-B says micro-averaging is used for all metrics, and the reported mIoU of 0.987 is therefore the global pixel IoU over 9 classes (8 fingertips + background). Because background dominates the images, this metric saturates near 1.0 even when per-class fingertip IoU is mediocre. The baseline numbers in Table III come from other papers with different protocols and likely different averaging schemes, so the comparison is apples-to-oranges. The authors disclose the micro-averaging, but they don't adjust the SOTA comparison for it.\n\nSecond, there are no error bars, no significance tests, and the test set is only 224 images from a private dataset. The paper doesn't describe whether the train/test split avoids overlapping subjects, which is critical in biometrics. Third, no code or data release; the baselines were not rerun on the same test set. These are all fixable.\n\nThe central method is sound and likely works as described on their data. The paper's contribution is real but narrow. The SOTA claim is overreaching as written.\n\nRecommendation: this deserves a serious referee, not a desk reject. A good referee could push the authors to rerun at least one baseline (e.g., DeepLabv3+ or U-Net) on their test set, report per-class mIoU alongside the micro-averaged number, and provide an identity-disjoint split. If they do that, the paper would be a solid applied contribution. As is, it's a conditional accept at best.","headline":"Sensible engineering result, but the SOTA claim rests on a micro-averaged mIoU that isn't comparable to the cited baselines.","tokens_in":12517,"tokens_out":1494,"would_cite":false,"duration_ms":16134,"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":"TipSegNet segments all four fingertips from one hand image with 0.999 accuracy.","keywords":["fingertip segmentation","contactless fingerprint","semantic segmentation","ResNeXt","Feature Pyramid Network","deep learning","biometrics","data augmentation"],"falsifier":"Run TipSegNet on an independently collected set of hand images with diverse backgrounds, lighting, and skin tones; if mIoU falls well below 0.987, the claimed generalization is not robust. Equally decisive: check whether the 224 test images contain any hands whose identities also appear among the 1,788 training images—overlap would inflate the reported metrics.","tokens_in":11601,"feed_emoji":"🖐️","tokens_out":4478,"duration_ms":38021,"temperature":0.7,"pith_summary":"TipSegNet is a deep learning model that segments all four fingertips directly from a single grayscale hand image, without a separate finger-detection step. The paper claims it reaches a mean Intersection over Union of 0.987 and an accuracy of 0.999 on a held-out test set, surpassing previously published whole-hand and single-finger segmentation methods. This matters because contactless fingerprint recognition depends on precise fingertip masks for pose correction and matching, and removing the detection stage simplifies the pipeline. The authors argue the gains come from a ResNeXt-101 backbone with a Feature Pyramid Network decoder plus heavy data augmentation.","feed_headline":"All four fingertips segmented in one pass at 0.999 accuracy","feed_subtitle":"TipSegNet skips finger detection and beats prior whole-hand and single-finger methods on contactless images.","key_machinery":"The load-bearing machinery is the combination of a ResNeXt-101 32×48d backbone, whose cardinality (32 parallel convolutional paths per block) captures richer feature interactions, with a Feature Pyramid Network that builds a multi-scale feature hierarchy through top-down pathways and lateral connections. The FPN produces independent predictions at each of the four backbone layers, upsamples them to a common resolution, sums them, and feeds the sum to a lightweight segmentation head. Transfer learning starts from weights pretrained on a large weakly supervised Instagram dataset, and training optimizes the Jaccard (IoU) loss. The design lets the network segment fingertips at varying scales and poses directly from the whole hand.","core_discovery":"The central claim is that a segmentation network built from a ResNeXt-101 32×48d encoder and a Feature Pyramid Network decoder, trained with aggressive geometric and photometric augmentation, can label nine classes—background plus the four fingertips of each hand—in whole-hand contactless images with near-saturated accuracy. On a test set of 224 images, TipSegNet attains mIoU 0.987, accuracy 0.999, F1 0.994, and recall 0.994, outperforming both two-class whole-hand segmentation approaches (Otsu, color histogram, Gaussian mixture, Mask R-CNN, DeepLabv3+) and single-finger methods (color-texture, mean shift, U-Net, EfficientNet, SqueezeNet). The paper also reports ablation studies showing that the ResNeXt-101 backbone gives only a small edge over ResNet-34/50/101 and that data augmentation barely changes the metrics, while raising training loss. The authors consequently note that the task is near saturation and that smaller backbones may be a more practical choice.","pith_inferences":["The very small gap between ResNet-34 and ResNeXt-101 suggests the 224-image test set may be too easy to discriminate model quality; a more challenging benchmark would likely produce a larger spread.","The model could be used to auto-label large contactless datasets, a direction the authors mention only as future work; that would let smaller student models train on far more data.","If augmented training barely changes test metrics but changes training loss, the augmentation may mainly act as a regularizer; testing on out-of-distribution images would reveal whether the claimed robustness actually comes from augmentation.","The reported accuracy of 0.999 should not be read as deployment-ready error-free segmentation; on video frames or unusual poses the error rate could be much higher."],"forward_implications":["Contactless fingerprint systems can drop the separate finger-detection stage, since TipSegNet outputs pixel masks for all four fingertips in one forward pass.","Downstream pose correction, unwarping, and minutiae matching receive cleaner fingertip regions, which should improve recognition accuracy.","The ablation results imply that a ResNet-34 or ResNet-50 backbone under the same FPN decoder achieves almost the same accuracy, so a practical deployment could use a far smaller model.","Because metrics are near saturation, the paper implies future progress needs harder datasets—more varied backgrounds, lighting, and skin tones—not bigger backbones."],"supporting_citations":[{"why":"Introduces the Feature Pyramid Network decoder that combines multi-scale features for the segmentation head.","marker":"[20]"},{"why":"Defines the ResNeXt architecture with cardinality, the backbone of TipSegNet.","marker":"[38]"},{"why":"Supplies the weakly supervised pretrained weights used to initialize the backbone via transfer learning.","marker":"[22]"},{"why":"Contributes 220 manually annotated hand images to the dataset and provides single-finger U-Net, EfficientNet, and SqueezeNet baselines.","marker":"[29]"},{"why":"Provides the DeepLabv3+ whole-hand fingertip segmentation baseline with feature-point construction.","marker":"[26]"},{"why":"Provides the Mask R-CNN, SegNet, HRNet, and DeepLab whole-hand baselines plus color and GMM methods.","marker":"[17]"},{"why":"Defines the Jaccard (IoU) loss used to train the segmentation model.","marker":"[36]"},{"why":"Underlies the U-Net architecture used as a single-finger baseline in the comparison.","marker":"[28]"}],"fun_headline_variants":["TipSegNet hits 0.999 accuracy for fingertip segmentation","Fingertip segmentation at 0.987 mIoU, 0.999 accuracy","One-pass fingertip segmentation beats prior methods on contactless images","Near-perfect fingertip segmentation in contactless hand imaging","Deep model segments fingertips at 0.999 accuracy in one pass"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported near-perfect scores assume the 224 test images are representative of real contactless captures and are correctly annotated, with no subjects shared between training and test sets.","fun_headline_variants_meta":{"raw":{"variants":["TipSegNet hits 0.999 accuracy for fingertip segmentation","Fingertip segmentation at 0.987 mIoU, 0.999 accuracy","One-pass fingertip segmentation beats prior methods on contactless images","Near-perfect fingertip segmentation in contactless hand imaging","Deep model segments fingertips at 0.999 accuracy in one pass"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000873,"raw_usage":{"total_tokens":3777,"prompt_tokens":940,"completion_tokens":2837,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":2743}},"tokens_in":556,"tokens_out":2837,"duration_ms":19853,"temperature":1.0,"reasoning_tokens":2743,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:18:07.686968+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TipSegNet on an independently collected set of hand images with diverse backgrounds, lighting, and skin tones; if mIoU falls well below 0.987, the claimed generalization is not robust. Equally decisive: check whether the 224 test images contain any hands whose identities also appear among the 1,788 training images—overlap would inflate the reported metrics.","supporting_citations":[{"cited_title":"Aggregated Residual Transformations for Deep Neural Networks","cited_arxiv_id":null,"evidence_quote":"Defines the ResNeXt architecture with cardinality, the backbone of TipSegNet."},{"cited_title":"Exploring the Limits of Weakly Supervised Pretraining, May","cited_arxiv_id":null,"evidence_quote":"Supplies the weakly supervised pretrained weights used to initialize the backbone via transfer learning."},{"cited_title":"Improving Sensor Interoperability between Contactless and Contact-Based Fingerprints Using Pose Correction and Unwarping","cited_arxiv_id":null,"evidence_quote":"Contributes 220 manually annotated hand images to the dataset and provides single-finger U-Net, EfficientNet, and SqueezeNet baselines."},{"cited_title":"Deep Learning-Based Semantic Segmentation for Touchless Fingerprint Recognition","cited_arxiv_id":null,"evidence_quote":"Provides the DeepLabv3+ whole-hand fingertip segmentation baseline with feature-point construction."},{"cited_title":"Towards Using Police Officers’ Business Smartphones for Con- tactless Fingerprint Acquisition and Enabling Fingerprint Comparison against Contact-Based Datasets","cited_arxiv_id":null,"evidence_quote":"Provides the Mask R-CNN, SegNet, HRNet, and DeepLab whole-hand baselines plus color and GMM methods."},{"cited_title":"Jaccard Metric Losses: Optimizing the Jaccard Index with Soft Labels","cited_arxiv_id":"2302.05666","evidence_quote":"Defines the Jaccard (IoU) loss used to train the segmentation model."},{"cited_title":"U-Net: Con- volutional Networks for Biomedical Image Segmentation","cited_arxiv_id":null,"evidence_quote":"Underlies the U-Net architecture used as a single-finger baseline in the comparison."}],"review_version":1}