{"id":"561466d2-9210-4cfa-b77b-6efa9dd4ac9b","arxiv_id":"2508.08094","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A multi-view computer vision pipeline detects and matches lateral roots in images, triangulates them, and refines the result with bundle adjustment to reconstruct 3D root skeletons from a few views, evaluated on a custom sweet potato dataset.","lead":"This paper builds a computer vision pipeline that turns a handful of photos of a plant root into a 3D skeleton, marking which parts are main versus lateral roots. It could help automate root trait measurement in plant breeding, but the tests use generated images of scanned roots rather than real field photos.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation uses rendered projections of manually denoised meshes with known camera poses, so the headline claim of extracting root architecture from a few real images is not yet supported.","rationale":"The reader's weakest assumption identifies the same linchpin: synthetic rendered inputs rather than real photographs. I agree with that diagnosis. The paper is internally coherent: detection losses, triangulation weighting, and SBA are standard formulations, and the skeleton-angle self-supervision is a reasonable idea. The central issue is not circularity but external validity. The evaluation protocol in Section V.A is a deliberate choice, acknowledged in the text as enabling quantitative comparison, but it means the reported numbers do not yet support the abstract's unqualified claim about real roots and breeding robots. The missing artifacts, undefined AP/AR matching criterion, and citation errors are secondary but compound the problem: even if the numbers were convincing, they cannot be checked or trusted without the underlying data and definitions. A real-image (or at least texture/pose-ablation) test is the one experiment that would discriminate between 'the method works on clean renderings' and 'the method works on actual root images.' Therefore the existing CONDITIONAL verdict remains appropriate; no change is needed.","tokens_in":10715,"tokens_out":6460,"duration_ms":76414,"concrete_test":"Acquire (or release) real multi-view photographs of 5–10 excavated root systems with ground-truth 3D skeletons from high-resolution CT or a scanner; run the full pipeline with camera poses recovered by COLMAP rather than taken from the renderer, and report 3D lateral-root AP/AR using the paper's matching criterion made explicit. If AP/AR remain near 0.77/0.58, the synthetic evaluation is a fair proxy; if they drop substantially (e.g., AP below 0.4), the abstract's general claim must be restricted to controlled rendered settings. As an immediate substitute if real data are unavailable, re-render the same meshes with constant albedo (no texture/shading) and randomly perturbed camera poses; a large performance drop would show that rendered texture and known poses are doing the work.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract, contribution 3) is that 3D root skeletons can be efficiently derived 'from a few images.' The only evidence for this, Section V.A, is a custom dataset of projection images rendered from 3D mesh models that were manually denoised from scanner point clouds. This removes precisely the properties that make real root imaging hard: real roots are near-textureless and monochromatic, but rendered meshes provide surface shading and edges that give the LightGlue matcher (Section III.B) usable keypoints; real roots are occluded by soil and clutter, while rendered images are clean; and real camera poses are unknown, while the virtual camera positions are known when rendering, so the triangulation in Section IV.A and the initial SBA pose estimates do not face the pose-estimation problem. The main-root connection algorithm in Section IV.C also assumes a single continuous foreground band in the image, an assumption that clean renderings satisfy but occluded field images do not. Consequently, the reported 3D AP/AR (0.77/0.58, Table III) and the comparison to AdaBins/MIM/Depth Anything measure performance on a synthetic-style proxy, not on the claimed application. The core triangulation and bundle-adjustment math is standard and internally coherent; the weakness is external validity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a pipeline for extracting 3D plant root skeletons from 3–10 multi-angle images. The method detects lateral roots with a YOLOv8-based detector that also predicts root keypoints, matches lateral roots across views using LightGlue keypoints and a voting scheme, triangulates the start/end points with weighted reprojection-error optimization, connects lateral roots to a main root using a growth-simulating row-propagation algorithm, and refines the skeleton and camera poses with a differentiable skeleton bundle adjustment (SBA) layer. Experiments on a custom dataset of 400 sweet-potato root mesh models, with rendered projection images and 40 test roots, report 2D AP/AR of 0.81/0.69 and 3D AP/AR of 0.77/0.58, outperforming AdaBins, MIM, and Depth Anything followed by Pc-Skeletor skeletonization. The paper claims to be the first to extract 3D root skeletons from images and to support automated breeding robots.","tokens_in":10923,"tokens_out":5547,"duration_ms":67350,"significance":"If the method works on real imagery, the contribution is potentially valuable for root phenotyping and breeding robotics. The modular design—detection, matching, triangulation, and SBA—is reasonable, and the idea of using skeletal-angle consistency for self-supervised bundle adjustment is interesting. The underlying triangulation and bundle-adjustment mathematics are standard and internally coherent. However, the only quantitative evaluation is performed on projection images rendered from manually denoised 3D meshes with known camera poses. This removes the key difficulties of real root imaging: occlusions, soil clutter, near-textureless surfaces, and unknown pose estimation. The reported AP/AR therefore measure performance on a clean synthetic-like proxy, not on the claimed application domain. The central claim is plausible but not yet supported; the paper needs either real-image validation or a substantial reduction in the scope of its claims. No code or dataset is released, and the evaluation protocol is not fully specified.","major_comments":[{"comment":"The load-bearing claim in the abstract and contribution 3—that the method extracts 3D root skeletons 'from a few images'—is evaluated only on projection images rendered from manually denoised 3D mesh models. The virtual camera positions are known when rendering, so the triangulation in Section IV.A and the SBA initialization do not face the pose-estimation problem that real multi-view root images would present. Moreover, rendered meshes provide shading and edge cues that real near-textureless roots lack, and the images are free of soil occlusion and clutter. Consequently, the AP/AR values in Table III do not support the stated real-world application. Please add experiments on real multi-view root images (including occlusions and estimated camera poses) or explicitly restrict the claims to clean rendered models.","section":"Section V.A, Table III"},{"comment":"The main-root connection algorithm assumes that in each image row the foreground is a single continuous interval [x1, x2]. This is true for the clean renderings used in Section V.A but is unlikely to hold for real root images with soil, clutter, crossed branches, or fragmented foreground regions. The mode-based propagation rule m = argmax_v |{M_(y-1)j = v}| depends critically on this continuity assumption. Since main-root skeleton extraction is an integral part of the method's central contribution, the algorithm should be tested on occluded or fragmented foregrounds, or modified to handle multiple intervals per row.","section":"Section IV.C, Eqs. (10)-(11)"},{"comment":"The precision and recall metrics for lateral-root detection and 3D skeleton extraction are not formally defined. It is unclear how predicted lateral roots or 3D skeletons are matched to ground-truth roots: is an IoU threshold, a point-to-mesh distance, or a graph-matching criterion used? Without this protocol, the reported numbers (e.g., 3D AP 0.77 and AR 0.58) cannot be independently interpreted or reproduced. Please specify the matching criterion, including thresholds and how multiple predictions per root are handled.","section":"Tables I-III, Section V.C"},{"comment":"The skeleton-angle self-supervision for the SBA layer is underspecified. The text says the loss is 'the difference between the original angle and the new angle after projection,' but it does not define what the 'original angle' is, when it is computed, or how 'new angle after projection' is obtained. If the original angle is measured from the current 3D skeleton estimate, the loss may act only as a regularizer rather than as a ground-truth-free self-supervision signal. Additionally, no ablation study isolates the contribution of the SBA layer to the final AP/AR, so the benefit of Eq. (9) is not demonstrated. Please provide a formal definition of the angle loss and add an SBA on/off comparison.","section":"Section IV.B"}],"minor_comments":[{"comment":"Typo: 'keleton extraction netwpork' should be 'skeleton extraction network'.","section":"Fig. 1 caption"},{"comment":"The y-coordinate keypoint scaling uses Aw_s (anchor width) for both b_y and b_x; presumably the y-coordinate should be scaled by Ah_s. Please correct.","section":"Eq. (2)"},{"comment":"The notation is inconsistent: Eq. (8) writes π(T_i, p_i) but the subscript in eg_i,j is j; it should likely be π(T_i, p_j). Please fix the index.","section":"Eq. (8)"},{"comment":"The sentence 'For lateral roots matched only in the second image' appears to mean the third image in the context of the pairwise triangulation; please clarify.","section":"Section IV.A"},{"comment":"Reference [2] is cited in the COLMAP discussion in Section II.B, but the work is about poultry farms and appears unrelated. Please verify the citation.","section":"References"},{"comment":"The comparison baselines are monocular depth estimators (AdaBins, MIM, Depth Anything) followed by Pc-Skeletor; classical multi-view stereo such as COLMAP is mentioned in the related work but not compared. Adding such a baseline would make the comparison more complete.","section":"Table III"},{"comment":"No code, trained models, or the custom dataset are released, which limits reproducibility; consider providing these or a detailed protocol.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is valid and is the primary reason for major revision. The paper's method may be sound for clean rendered inputs, but the abstract and conclusion overstate the support for real-world root imaging. If the authors can supply real multi-view root experiments (even a small set) or clearly scope all claims to the rendered domain, the contribution could be acceptable. No issues of citation integrity were apparent beyond the irrelevant reference [2]."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the method is a sensible integration of off-the-shelf pieces, and the SBA layer with learned damping and angle-loss is a small original twist. But the only quantitative evidence comes from rendered images of manually cleaned meshes, not real root photographs. So the abstract's 'from a few images' claim should read 'from a few clean projections.'\n\nThe pipeline is clearly described: YOLOv8 detects lateral root boxes with start/end keypoints, LightGlue gives matches, a voting matrix pairs boxes, weighted triangulation from two view-pairs reduces error, and SBA with a learned damping factor and self-supervised angle-consistency loss refines poses and points. The main-root connection via row-wise propagation is a neat way to chain lateral roots. The math in Sections IV.A and IV.B is standard triangulation/BA and internally consistent. On their 40-root synthetic test set, the method beats depth-estimation baselines by a wide margin (3D AP 0.77 vs 0.37 for Depth Anything).\n\nThe problem is external validity. The test images are renderings of meshes that were manually denoised from scanner point clouds, with known virtual camera poses. That removes texturelessness, occlusion, and pose estimation—the very things that make real root imaging hard. The main-root connection assumes a clean continuous foreground band; a field image with soil and overlapping roots would break it. No code, data, or model is released, and the thresholds and loss weights are unstated, so replication is not possible. Metrics are point estimates without error bars or ablations; the SBA layer appears to be doing useful work, but we don't know how much. There is also a clear citation error: reference [2] is a poultry-farm paper, not COLMAP. These are fixable, but they mean the current evidence supports a narrow synthetic regime, not the claimed application.\n\nPlant phenotyping researchers will find the skeletonization and topology ideas useful, and CV people working on thin-structure reconstruction might learn from the SBA scheme. But don't cite it yet.\n\nIt deserves a serious referee, not a desk reject, because the core method is coherent and the synthetic evaluation is a reasonable first step. But the referee should demand real-image tests (even lab photos), ablations, error bars, and code/artifacts. With those, the paper could be solid. As it stands, it is a promising technique with a validation gap.","headline":"Promising pipeline for 3D root skeletons, but the evaluation is on clean renders, not real photos, so the headline claim overreaches.","tokens_in":11503,"tokens_out":4000,"would_cite":false,"duration_ms":44653,"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":"This paper claims that 3D root skeletons can be reconstructed from as few as three multi-angle images by detecting, matching, and triangulating lateral roots, and that the recovered architecture is close to ground truth.","keywords":["3D root skeleton","plant phenotyping","lateral root detection","multi-view triangulation","bundle adjustment","sweet potato roots","breeding robots"],"falsifier":"Photograph real root systems with a handheld camera and unknown poses, or render the same models with soil occlusion and perturbed cameras, then compare recovered skeletons against CT-derived ground truth. If lateral-root precision and recall fall to depth-estimation baseline levels, the clean-mesh known-pose rendering is what produces the reported accuracy.","tokens_in":10527,"feed_emoji":"🌱","tokens_out":8466,"duration_ms":81996,"temperature":0.7,"pith_summary":"This paper aims to establish that dense, texture-poor plant root systems can be converted into 3D skeletons from as few as three to ten ordinary images, without depth sensors or synchronized camera rigs. The key move is to detect, match, and triangulate whole lateral roots instead of reconstructing point clouds: a convolutional detector finds lateral roots and their endpoints, keypoint votes link the same root across views, and weighted triangulation turns matched roots into 3D lines. A differentiable bundle-adjustment layer then refines skeleton points and camera poses, supervised by the fact that angles between roots are invariant in 3D. The main root is assembled by connecting lateral-root starting points in growth order, which also separates the two root types. The pipeline is validated on a custom set of 400 scanned sweet-potato root meshes, with 40 held out for testing, reporting higher 3D lateral-root precision and recall than depth-estimation-plus-skeletonization baselines.","feed_headline":"A few photos reconstruct 3D root skeletons","feed_subtitle":"Detection, matching, and triangulation recover branching architecture that depth-based pipelines miss.","key_machinery":"The load-bearing object is the lateral root treated as a detected and matched unit rather than a set of independent points. A convolutional detector outputs boxes plus start/end keypoints for each lateral root; a keypoint-voting step builds a match-score matrix to associate boxes across views; weighted triangulation fuses two-view reconstructions; and a differentiable Skeleton Bundle Adjustment layer minimizes reprojection error and angle discrepancies, predicting the Levenberg-Marquardt damping factor with attention so the whole optimization is trainable by backpropagation. The main root is assembled by a row-wise growth-simulation matrix that connects lateral-root origins in order.","core_discovery":"The paper's central claim is that 3D root skeletons can be reconstructed from as few as three to ten images by detecting, matching, and triangulating lateral roots instead of dense point-cloud reconstruction. Lateral roots are located as objects with start and end keypoints; matching across views is decided by keypoint votes between bounding boxes; two-way triangulation is weighted by reprojection error; and a learnable bundle-adjustment layer, supervised by the invariance of root angles, refines skeleton and camera poses. Main-root topology is obtained by connecting lateral-root origins in growth order. On 40 held-out sweet-potato root systems, the reported 3D lateral-root precision is 0.77","pith_inferences":["The lateral-root-as-object strategy should transfer to other self-occluding, texture-poor branching structures — tree canopies, vasculature, neurons — wherever branches can be defined as detections.","The clean-mesh testbed with known poses is the main gap between these numbers and field performance; the decisive next test is unknown-pose photography of real soil-grown roots against CT or MRI ground truth.","The growth-simulation rule for connecting lateral roots assumes downward, tip-first growth; root systems with horizontal or upward-growing branches would need a learned connectivity head.","If pose-noise robustness holds, the pipeline could run on video from a single moving camera, making 3D root phenotyping possible in the field rather than the lab."],"forward_implications":["Root architecture phenotyping — root counts, branching order, angles, lengths — becomes possible from a few handheld or robot-arm images instead of multi-camera or destructive measurements.","Primary and lateral roots are separated automatically, so trait measurements can be reported per root type.","The angle-consistency self-supervision removes the need for dense 3D ground truth during skeleton refinement.","With three views enough for reconstruction, capture can be done by a moving camera, simplifying breeding-robot integration.","Breeding robots could use the resulting skeleton as direct input for selecting plants with favorable root architecture."],"supporting_citations":[{"why":"Supplies the convolutional object-detection backbone used to localize lateral roots and their start/end keypoints.","marker":"[12]"},{"why":"Provides the local feature extraction and matching whose keypoints vote to associate detected lateral roots across views.","marker":"[16]"},{"why":"Serves as the supervised depth-estimation baseline whose point clouds are skeletonized for comparison.","marker":"[1]"},{"why":"Serves as a masked-image-modeling depth baseline in the multi-view skeleton comparison.","marker":"[29]"},{"why":"Serves as the large-scale monocular depth baseline in the multi-view skeleton comparison.","marker":"[32]"},{"why":"Provides the point-cloud skeletonization baseline applied to depth-derived point clouds.","marker":"[22]"},{"why":"Supplies the differentiable bundle-adjustment formulation that the paper's skeleton bundle-adjustment layer adapts.","marker":"[26]"},{"why":"Serves as the direct keypoint-detection baseline for lateral-root matching comparisons.","marker":"[6]"}],"fun_headline_variants":["3D root skeletons from a handful of photos","Few images, full 3D root skeleton extraction","Root architecture in 3D from sparse views","Triangulation yields 3D root skeletons from few images","From snapshots to 3D root structure"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that projection images rendered from manually denoised 3D meshes, with camera poses known exactly, stand in for real photographs of roots growing in soil.","fun_headline_variants_meta":{"raw":{"variants":["3D root skeletons from a handful of photos","Few images, full 3D root skeleton extraction","Root architecture in 3D from sparse views","Triangulation yields 3D root skeletons from few images","From snapshots to 3D root structure"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000358,"raw_usage":{"total_tokens":1790,"prompt_tokens":770,"completion_tokens":1020,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":945}},"tokens_in":514,"tokens_out":1020,"duration_ms":11196,"temperature":1.0,"reasoning_tokens":945,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T21:38:55.967288+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Photograph real root systems with a handheld camera and unknown poses, or render the same models with soil occlusion and perturbed cameras, then compare recovered skeletons against CT-derived ground truth. If lateral-root precision and recall fall to depth-estimation baseline levels, the clean-mesh known-pose rendering is what produces the reported accuracy.","supporting_citations":[{"cited_title":"Ultralytics YOLO, Jan","cited_arxiv_id":null,"evidence_quote":"Supplies the convolutional object-detection backbone used to localize lateral roots and their start/end keypoints."},{"cited_title":"Light- glue: Local feature matching at light speed","cited_arxiv_id":null,"evidence_quote":"Provides the local feature extraction and matching whose keypoints vote to associate detected lateral roots across views."},{"cited_title":"Adabins: Depth estimation using adaptive bins","cited_arxiv_id":null,"evidence_quote":"Serves as the supervised depth-estimation baseline whose point clouds are skeletonized for comparison."},{"cited_title":"Revealing the dark secrets of masked image modeling","cited_arxiv_id":null,"evidence_quote":"Serves as a masked-image-modeling depth baseline in the multi-view skeleton comparison."},{"cited_title":"Cherrypicker: Semantic skeletonization and topological reconstruction of cherry trees","cited_arxiv_id":null,"evidence_quote":"Provides the point-cloud skeletonization baseline applied to depth-derived point clouds."},{"cited_title":"Su- perpoint: Self-supervised interest point detection and description","cited_arxiv_id":null,"evidence_quote":"Serves as the direct keypoint-detection baseline for lateral-root matching comparisons."}],"review_version":1}