{"id":"8a505a91-c5d5-428b-918c-a1e2bc67626a","arxiv_id":"2411.13544","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"DIS-Mine fuses enhanced SAM and Mask R-CNN masks to segment objects in dark mine images, reporting 86% F1 on one external benchmark but only 70.2% on its own new ImageMine dataset.","lead":"This paper presents DIS-Mine, a system that finds and outlines objects in near-dark underground mine images by combining two existing segmentation models with a low-light enhancement step. It also introduces ImageMine, a new set of dark mine images, and claims large performance gains that the paper's own tables do not fully support.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim depends on ImageMine evaluation masks that appear to be generated by the same SAM model DIS-Mine uses; without an independent human-labeled test split, DIS-Mine's advantage over non-SAM baselines may be an artifact of shared label generation.","rationale":"The reader's REJECT is well-supported by the text. My independent reading converges on the same load-bearing point: the ImageMine evaluation labels may not be independent of the method, because the SAM-based auto-annotation pipeline described in Sections III.C and V.A appears to generate the masks used for both training and evaluation, and SAM is a component of DIS-Mine. This is not a matter of style or of disagreement with the field's consensus; it is a correctness risk in the central empirical claim. The abstract's headline numbers are also drawn from the DsLMF+ row of Table II rather than the new ImageMine dataset, and the dramatic 15x/80% claims are unsupported by any table or experiment. A human-labeled holdout split and a recomputation of Table II on that split would settle whether the claimed advantage is real. Since this concern reinforces rather than changes the reader's verdict, I recommend no change to the REJECT decision.","tokens_in":12484,"tokens_out":3227,"duration_ms":33071,"concrete_test":"Obtain or construct an independent human-annotated test set for ImageMine (e.g., re-annotate a random sample of 100 held-out images with two or more annotators). Compute human-vs-SAM mask agreement (IoU); then recompute the Table II ImageMine F1/mIoU for DIS-Mine and all baselines using only human labels, with the same matching threshold. If DIS-Mine's margin over SAM/Mask2Former disappears or reverses, the central superiority claim is an artifact of shared label generation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's core claim—that DIS-Mine outperforms state-of-the-art instance segmentation in poor-light mine scenes—is only meaningful if the evaluation ground truth is independent of the method. Section III.C states that SAM was fine-tuned on 510 manually annotated images and then used to auto-annotate the remainder of ImageMine. Section V.A then says these SAM-generated image-mask pairs were used to train the Mask R-CNN component, and SAM itself was retrained on both manually and automatically annotated data. No separate human-labeled test split, no human/SAM agreement measure, and no exclusion of auto-annotated images from evaluation are reported. If the Table II ImageMine masks come from the same SAM pipeline, DIS-Mine—which explicitly fuses SAM with Mask R-CNN—is being scored partly against its own label generator, while baselines such as Mask R-CNN or Mask2Former are scored against a SAM-specific label distribution. This inflates the apparent margin and undermines the headline. Independently, the abstract's 86.0/72.0 numbers are only the DsLMF+ row of Table II; ImageMine gives 70.2/60.5, and the claimed 'at least 15x improvement and up to 80% higher precision' does not appear in any table.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DIS-Mine, an instance-segmentation pipeline for underground mine images under poor-light or near-dark conditions. The pipeline consists of four components: a KinD-based brightness-enhancement network, a fine-tuned Segment Anything Model (SAM) that takes Mask R-CNN predictions as prompts, a Mask R-CNN variant with weighted Dice plus focal mask loss, and a mask-alignment stage that uses ORB feature matching, morphological operations, and a rule-based 4x4 grid positional classifier. The authors also introduce ImageMine, a low-light underground mine dataset with a subset of manually annotated images, and describe an automatic SAM-based annotation pipeline for the rest. Experiments compare DIS-Mine against SAM, Mask R-CNN, Mask2Former, and ISD on ImageMine, LIS, and DsLMF+, reporting F1 and mIoU. The paper claims superior performance, including an 86.0% F1 score and 72.0% mIoU, with 'at least 15x improvement' over state-of-the-art methods.","tokens_in":12657,"tokens_out":5504,"duration_ms":60147,"significance":"If the ImageMine results are validated by an independent label protocol, the dataset itself would be a useful contribution to disaster-response vision in underground mines, and the proposed SAM-plus-Mask R-CNN fusion with low-light enhancement is a reasonable engineering direction. The paper also makes the dataset publicly available, which is a strength. However, the central quantitative claims are not currently supported by the reported experiments: the headline numbers apply only to the DsLMF+ benchmark, and the ImageMine evaluation appears to depend on labels generated by the same SAM model that the proposed method integrates. Because the comparison against baselines is the core of the paper, this evaluation gap is load-bearing.","major_comments":[{"comment":"The abstract states that DIS-Mine 'achieves a superior F1 score of 86.0% and mIoU of 72.0%, outperforming state-of-the-art instance segmentation methods, with at least 15x improvement and up to 80% higher precision in object detection.' This is not supported by the paper's own Table II. The 86.0% F1 and 72.0% mIoU are the DsLMF+ results; on ImageMine, the dataset introduced by the paper, DIS-Mine obtains 70.2% F1 and 60.5% mIoU. No table or figure anywhere reports a '15x improvement' or an '80% higher precision' figure, and object-detection precision is not an evaluated metric in Section V. Please restate the abstract using per-dataset results and remove claims that no experiment supports.","section":"Abstract and Table II"},{"comment":"The ImageMine evaluation may be circular. Section III.C states that SAM was fine-tuned on 510 manually annotated images and then used to auto-annotate the remainder of ImageMine. Section V.A states that these SAM-generated image-mask pairs were used to train the Mask R-CNN component and that SAM was retrained on both manually and automatically annotated data. Since DIS-Mine explicitly fuses SAM with Mask R-CNN, its outputs share the same label distribution that produces the ground-truth masks, whereas the baselines are evaluated against masks produced by a generator they do not contain. The paper does not describe a separate human-labeled test split, a human/SAM agreement measure, or any exclusion of auto-annotated images from evaluation. Without an independent test set that is not used in SAM fine-tuning or Mask R-CNN training, the reported advantage over non-SAM baselines on ImageMine is not trustworthy. A revision must use a held-out manual test set and report agreement between SAM auto-annotations and human labels.","section":"Section III.C and Section V.A"},{"comment":"The experimental protocol is underspecified in ways that affect the interpretation of the reported margins. The paper does not state the total number of images in ImageMine, the number of auto-annotated images, or the train/validation/test split for any dataset. It also does not define the matching threshold used to compute instance-level F1 and mIoU, nor say whether metrics are averaged over multiple runs. Without these details, the differences between DIS-Mine and the second-best models on ImageMine (70.2% versus 68.7% F1; 60.5% versus 60.0% mIoU) cannot be assessed for significance or reproducibility. Please provide the full protocol, including dataset statistics and evaluation thresholds.","section":"Section V.A and Section V.C"},{"comment":"No ablation study is reported, although the paper's claimed innovations are the four components of DIS-Mine. In particular, the rule-based 4x4 grid assignment in Section IV.D, which classifies objects as road/wall/roof according to their position in the image, is a strong inductive bias that could inflate performance on geometrically similar classes. The paper should add ablations that remove or deactivate (i) the brightness-improvement component, (ii) the SAM integration, (iii) the enhanced mask loss, and (iv) the positional grid rule, and report each variant's F1 and mIoU on ImageMine. This is necessary to determine which component is responsible for the reported gains.","section":"Section IV.D and Section V.C"}],"minor_comments":[{"comment":"The paper is inconsistent about the number of annotated classes: the Introduction says 'manually annotated a subset of images into five classes' and then later lists six categories (road, wall, roof, people, equipment, corridor). Please correct this inconsistency.","section":"Abstract and Introduction"},{"comment":"The phrase 'about 100-80 images were extracted' should presumably read '80-100 images'; please fix the range.","section":"Section III.A"},{"comment":"Algorithm 1's caption says 'offers a explanation'; this should be 'offers an explanation'.","section":"Section IV.A"},{"comment":"The acronym ORB is expanded as 'oriented fAST and rotated brief'; the correct capitalization is 'Oriented FAST and Rotated BRIEF'.","section":"Section IV.D"},{"comment":"The text says Mask2Former 'consistently ranks as the second-best model' after DIS-Mine, but in the LIS dataset row of Table II, ISD reports a higher F1 score (61.7%) than Mask2Former (62.0% is actually higher, so check: ISD 61.7%, Mask2Former 62.0%; the statement is true for F1, but for mIoU ISD is 49.8% versus Mask2Former 45.8%, so the sentence should be qualified by metric). Please make the ranking statement consistent with Table II.","section":"Section V.C"},{"comment":"For the People class, F1 and IoU are both reported as 72.6%; this coincidence is suspicious and should be checked, as F1 and IoU are different quantities.","section":"Table III"}],"recommendation":"major_revision","confidential_remarks":"I considered reject, because the central ImageMine comparison may be invalidated by label circularity and the abstract contains unsupported quantitative claims. I recommend major revision rather than immediate rejection because the evaluation flaw is potentially fixable: the authors already have 510 manually annotated images and could hold out a human-labeled test set that is not used in SAM fine-tuning or Mask R-CNN training, then re-run the comparisons and correct the abstract. If such a test set cannot be supplied, rejection would be appropriate. The public release of ImageMine is a positive factor, but the method's claimed advantage over baselines must be demonstrated on independent labels."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main thing to know: this paper's real value is the ImageMine dataset, a genuinely useful collection of low-light underground mine images. The pipeline itself is an integration of existing pieces—KinD for enhancement, SAM with Mask R-CNN prompts, ORB-based mask fusion, and a weighted dice plus focal loss. That's a reasonable engineering contribution, and on the DsLMF+ benchmark the fusion does beat each component alone by a few points. Credit where it's due: they collected and manually annotated a subset of the data, describe the acquisition setup, and compare against sensible baselines.\n\nBut the paper has a load-bearing evaluation problem, and the stress-test note is right about it. Section III.C says SAM was fine-tuned on 510 manual labels and then used to auto-annotate the rest of ImageMine. Section V.A says those auto-generated masks trained Mask R-CNN, and SAM itself was retrained on both manual and automatic labels. There is no separate human-labeled test split, no SAM-vs-human agreement measure, and no guarantee the evaluation masks in Table II are independent of the SAM pipeline. Since DIS-Mine explicitly fuses SAM with Mask R-CNN, it is being scored partly against labels produced by its own component, while baselines like Mask R-CNN and Mask2Former are scored against a SAM-specific label distribution. That inflates the margin and makes the ImageMine comparison uninterpretable.\n\nOn top of that, the abstract's central quantitative claims are not supported by the paper's own data. The 86.0% F1 and 72.0% mIoU are the DsLMF+ row of Table II, not ImageMine (70.2/60.5), and the 'at least 15x improvement and up to 80% higher precision' appears nowhere in the results. The evaluation is also under-specified: no train/val/test split, no matching threshold for F1/IoU, no error bars, and no significance tests. The post-hoc merging of road, roof, and wall into 'surrounding' is another protocol shift that needs to be flagged. Dataset link and code are absent.\n\nIn summary: the dataset is worth something, and a modest claim about the fusion beating individual components on DsLMF+ is supported. But the headline claims and the ImageMine evaluation as presented are not. The paper deserves a serious referee because the dataset and the pipeline question are legitimate, but it needs major revision: an independent human-labeled test split, a corrected abstract, and a proper protocol description. I'd send it to review with a clear request for those changes.\n\nRecommendation: engage with it, but reject in its current form.","headline":"The ImageMine dataset is a real asset, but the headline numbers are overclaimed and the evaluation on ImageMine is undermined by auto-annotated labels from the same SAM model that DIS-Mine uses.","tokens_in":13347,"tokens_out":1876,"would_cite":false,"duration_ms":21208,"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":"DIS-Mine fuses SAM and Mask R-CNN to segment disaster-affected underground mine scenes in poor light, reporting 86.0% F1 and 72.0% mIoU on DsLMF+ and 70.2%/60.5% on its new ImageMine dataset.","keywords":["instance segmentation","low-light image enhancement","underground mining safety","disaster response","Segment Anything Model","Mask R-CNN","image dataset","mask alignment"],"falsifier":"Take a hold-out set of dark mine images, label every mask by hand without any SAM assistance, then evaluate DIS-Mine and the baselines on that split; if DIS-Mine's F1/mIoU advantage over SAM and Mask R-CNN collapses, the reported superiority is an artifact of sharing the SAM label generator rather than a genuine property of the fusion.","tokens_in":12124,"feed_emoji":"⛏️","tokens_out":5906,"duration_ms":56584,"temperature":0.7,"pith_summary":"DIS-Mine tries to establish that instance segmentation in near-dark underground mines can be made reliable enough for rescue use, despite the noise, color distortion, and low contrast that defeat standard models. The method couples a KinD-based brightness-enhancement network with two complementary segmenters—a fine-tuned Segment Anything Model and a Mask R-CNN—and fuses their masks through ORB feature matching, intersection, and morphological cleanup. The authors also introduce ImageMine, a real-world dataset of dark experimental-mine footage labeled for people, equipment, corridors, and surrounding structures. On DsLMF+ they report 86.0% F1 and 72.0% mIoU, beating SAM, Mask R-CNN, and Mask2Former, and on ImageMine they report 70.2% F1 and 60.5% mIoU. If the method works as claimed, first responders could receive instance-level maps of people and hazards in otherwise invisible mine conditions.","feed_headline":"Two-model fusion beats single models in dark mine segmentation","feed_subtitle":"DIS-Mine reports 86% F1 on DsLMF+ and 70% on the new ImageMine set, edging SAM, Mask R-CNN, and Mask2Former.","key_machinery":"The load-bearing mechanism is the mask-alignment fusion of two heterogeneous segmenters. A KinD (Retinex-based) network first brightens the dark image. A fine-tuned SAM takes the enhanced image plus bounding-box and class prompts generated by Mask R-CNN and produces one mask; Mask R-CNN, trained with a weighted-Dice-plus-focal-loss mask term, produces another mask. ORB feature matching aligns the two masks, their intersection is taken, and dilation/erosion plus a 4x4-grid rule (e.g., anything in the lower row is labeled road) yield the final instance mask. The idea is that the two models' errors are different enough that intersecting their aligned outputs removes false positives while the feature-matching step fixes spatial misalignment.","core_discovery":"The paper's central claim is that its four-component pipeline—image brightness improvement, SAM-based instance segmentation with prompts from Mask R-CNN, Mask R-CNN with a weighted-Dice-plus-focal mask loss, and final mask alignment by ORB feature matching—outperforms each component alone and every single-model state-of-the-art baseline under poor-light conditions. The reported evidence is Table II: on DsLMF+ DIS-Mine reaches 86.0% F1/72.0% mIoU (SAM: 84.0/71.0; Mask R-CNN: 80.0/68.0; Mask2Former: 83.0/72.0); on ImageMine it reaches 70.2/60.5 (SAM: 68.7/60.0; Mask R-CNN: 65.0/56.0; Mask2Former: 67.2/58.0); and on LIS it reaches 63.2/47.0, above ISD's 61.7/49.8 on F1 but below on mIoU. The abstract's additional statement of 'at least 15x improvement and up to 80% higher precision in object detection' is not backed by a comparison table in the manuscript.","pith_inferences":["If the fusion advantage is real, a natural deployment is untethered rescue robots: DIS-Mine's pipeline could run on the enhanced image stream and broadcast instance masks of people and corridors to surface teams; the paper does not test this.","A direct test of the mask-alignment step would be to ablate the ORB intersection and keep either the SAM or Mask R-CNN branch; the paper reports no per-component ablation, so the contribution of alignment to the 70.2/60.5 scores is untested.","Because the auto-annotation pipeline uses SAM, an independent human-labeled test split would separate the method's true generalization from the label generator's familiarity; the paper does not provide one.","The stated future direction of fusing thermal or LiDAR modalities is a natural next step; in total darkness, active sensors would supply geometric cues that RGB enhancement cannot recover."],"forward_implications":["Deployed on mine surveillance feeds, DIS-Mine could give rescue teams near-real-time instance masks of people, equipment, and corridors in darkness.","The automatic SAM-based annotation pipeline can label new images from similar underground environments, reducing the manual labeling burden from thousands of images to a few hundred.","Fusing two segmenters with alignment beats each segmenter alone on all three datasets reported, suggesting the fusion strategy, not just better features, drives the gains.","On the LIS benchmark DIS-Mine's F1 (63.2%) tops the specialized low-light segmenter ISD (61.7%), indicating the method transfers beyond mines.","The reported margins on DsLMF+ are small (1–6 points over SAM), so the practical benefit is a moderate, consistent improvement rather than a step change."],"supporting_citations":[{"why":"SAM is the large pretrained segmenter that DIS-Mine fine-tunes and uses as one of the two mask generators.","marker":"[23]"},{"why":"Mask R-CNN is the other mask generator, whose mask loss DIS-Mine modifies to weighted Dice plus focal.","marker":"[14]"},{"why":"KinD is the Retinex-based low-light enhancement network that brightens images before both segmenters.","marker":"[12]"},{"why":"DsLMF+ is the underground-longwall dataset where DIS-Mine reports its best numbers (86.0/72.0).","marker":"[27]"},{"why":"ISD and the LIS dataset supply the specialized low-light segmentation baseline and benchmark the paper compares against.","marker":"[10]"},{"why":"ORB is the feature-matching algorithm used to align SAM and Mask R-CNN masks before intersection.","marker":"[26]"},{"why":"Mask2Former is the universal-segmentation baseline used in Table II.","marker":"[28]"},{"why":"VGG Image Annotator is the tool used for the 510 manually annotated images that seed the SAM auto-annotation pipeline.","marker":"[24]"}],"fun_headline_variants":["DIS-Mine fuses SAM and Mask R-CNN for dark mine rescue","New pipeline beats single models in poor-light underground mine","86% F1 in darkness: DIS-Mine segment disaster zones","Four-component fusion improves mine disaster detection","New ImageMine dataset tests DIS-Mine in real mines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the ImageMine ground-truth masks used for training and evaluation are accurate and independent of the SAM model embedded in DIS-Mine, even though Section III.C uses a fine-tuned SAM to auto-annotate the dataset and Section V.A trains Mask R-CNN on those auto-generated masks.","fun_headline_variants_meta":{"raw":{"variants":["DIS-Mine fuses SAM and Mask R-CNN for dark mine rescue","New pipeline beats single models in poor-light underground mine","86% F1 in darkness: DIS-Mine segment disaster zones","Four-component fusion improves mine disaster detection","New ImageMine dataset tests DIS-Mine in real mines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000801,"raw_usage":{"total_tokens":3608,"prompt_tokens":1120,"completion_tokens":2488,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":736,"completion_tokens_details":{"reasoning_tokens":2405}},"tokens_in":736,"tokens_out":2488,"duration_ms":20241,"temperature":1.0,"reasoning_tokens":2405,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:18:22.276246+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a hold-out set of dark mine images, label every mask by hand without any SAM assistance, then evaluate DIS-Mine and the baselines on that split; if DIS-Mine's F1/mIoU advantage over SAM and Mask R-CNN collapses, the reported superiority is an artifact of sharing the SAM label generator rather than a genuine property of the fusion.","supporting_citations":[{"cited_title":"Segment anything,","cited_arxiv_id":null,"evidence_quote":"SAM is the large pretrained segmenter that DIS-Mine fine-tunes and uses as one of the two mask generators."},{"cited_title":"Kindling the darkness: A practical low-light image enhancer,","cited_arxiv_id":null,"evidence_quote":"KinD is the Retinex-based low-light enhancement network that brightens images before both segmenters."},{"cited_title":"Dslmf+: An open dataset for intelligent recognition of abnormal condition in underground longwall mining face,","cited_arxiv_id":null,"evidence_quote":"DsLMF+ is the underground-longwall dataset where DIS-Mine reports its best numbers (86.0/72.0)."},{"cited_title":"Instance segmentation in the dark,","cited_arxiv_id":null,"evidence_quote":"ISD and the LIS dataset supply the specialized low-light segmentation baseline and benchmark the paper compares against."},{"cited_title":"Masked-attention mask transformer for universal image segmentation,","cited_arxiv_id":null,"evidence_quote":"Mask2Former is the universal-segmentation baseline used in Table II."},{"cited_title":"VGG image annotator (VIA),","cited_arxiv_id":null,"evidence_quote":"VGG Image Annotator is the tool used for the 510 manually annotated images that seed the SAM auto-annotation pipeline."}],"review_version":1}