{"id":"da7e549e-4b38-40f8-8656-c6616ff900ee","arxiv_id":"2509.08333","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Self-supervised fine-tuning of SuperPoint using visual odometry reprojection errors as labels produces more uniform features and improves trajectory estimates on the Pohang dataset, according to qualitative results only.","lead":"A team fine-tunes the SuperPoint feature extractor using feedback from a visual odometry system, so the detector learns to ignore unstable areas like water or terrain. Early results on a maritime dataset show better feature spreading, but no quantitative accuracy numbers are reported.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The self-supervised labels come from the VO's own pose estimate, creating a circular signal that could reinforce the VO's bias in the very low-texture/dynamic scenes the method targets.","rationale":"The reader's weakest assumption identifies the VO-pose-based labeling as the key fragility. I agree: the method explicitly uses the VO's own pose estimate to decide which features are 'good' via reprojection error. If that pose is systematically wrong in the low-texture, dynamic, or out-of-distribution scenes that motivate the paper, the self-supervision will reward features that are consistent with the wrong pose, reinforcing the very errors the method claims to correct. This is a classic confirmation-bias loop. The paper's qualitative evidence (Figs. 3-5) cannot rule this out because there is no quantitative trajectory error, no ground-truth comparison, and no disclosed train/test split. The SuperGlue descriptor mismatch is also a concern, but the core validity of the training paradigm hinges on the label quality. If the circularity test I propose shows high label agreement, the concern is resolved; if not, the central claim is unsupported or even harmful. Thus, the verdict remains CONDITIONAL: the paper should be accepted only if the authors can validate the labels or show that the feedback loop does not dominate. The reader's low confidence is justified, and my analysis does not change that.","tokens_in":4340,"tokens_out":6098,"duration_ms":61842,"concrete_test":"On a benchmark with ground truth trajectory (e.g., KITTI odometry), for each stereo pair in a challenging sequence, detect and match features, and compute reprojection errors using (a) the VO-estimated relative pose (as in the paper) and (b) the ground-truth relative pose. Count the fraction of features labeled 'good' (e.g., error below threshold) under (a) that are also 'good' under (b). If this agreement is below ~0.8, or if the VO-pose labels systematically favor features that later correlate with large ATE, the training signal is unreliable. A stronger test: train two SuperPoint models, one with VO-pose labels and one with ground-truth-pose labels, and compare final VO trajectory accuracy; if the VO-pose-trained model does not beat the pre-trained baseline, the circularity is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the self-supervised loss (Eq. 1) identifies 'good' features from low reprojection error, but those reprojection errors are computed using the relative pose estimated by RTAB-Map from the *same* features being trained. This creates a circular dependency: if the VO's pose estimate is biased—e.g., tracking moving water reflections, low-texture terrain, or other unreliable regions—those features will appear geometrically consistent under the wrong pose and be scored as 'good' by Li(X,Y). The fine-tuning may then amplify exactly the biases the method intends to remove. The paper provides no evidence that the VO pose estimates on Pohang are accurate enough for the labels to be trustworthy; the trajectory plots (Fig. 5) show no ground truth, so the reader cannot judge which trajectory is correct. This is not just a missing metric: it is a potential invalidation of the training signal in the very scenes where the method aims to help.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a self-supervised fine-tuning procedure for the SuperPoint feature extractor within an RTAB-Map visual odometry pipeline, with SuperGlue as the matcher. The key idea is to use the VO's own estimated relative pose and the resulting reprojection errors to label image-region-specific 'good' features, and then train SuperPoint with a weighted loss (Eq. 1) that combines keypoint losses on original and warped images, a peaky loss, and a descriptor hinge loss. The authors present qualitative feature maps on Pohang and KITTI data and trajectory plots on Pohang sequences, claiming that the self-supervised model produces more uniformly distributed features on rigid structures while suppressing unreliable regions such as water and terrain, and that this improves trajectory estimation in narrow and offshore scenes. The paper explicitly states that the work is at an early stage, that SuperGlue was not retrained, and that the current results are 'not yet sufficient' to fully answer what defines good deep features to track.","tokens_in":4645,"tokens_out":3681,"duration_ms":47369,"significance":"If validated, the proposed self-supervised adaptation is practically valuable: it could adapt generic deep feature extractors to a target deployment environment without manual annotation, using only geometric consistency signals derived from an existing VO/SLAM system. The idea of task-specific feedback for feature learning is interesting, and the use of a real maritime dataset (Pohang) strengthens the motivation. The paper also has the virtue of making a concrete falsifiable claim: fine-tuning SuperPoint with the proposed loss changes feature distribution and improves VO trajectory quality. However, the current evidence is almost entirely qualitative, and the self-supervision signal is derived from the very VO system the method is meant to improve, creating a circularity risk that is not addressed experimentally. The absence of quantitative metrics, training details, and a controlled study of the label-generation mechanism means the central claim is not yet established.","major_comments":[{"comment":"The central claim that the self-supervised SuperPoint 'outperforms' the other VO variants is not supported by any quantitative evidence. The trajectory plots show no ground truth, and no ATE, RPE, scale-normalized error, or statistics over repeated runs are reported. Without such metrics, the reader cannot assess which trajectory is more accurate, especially since differences appear mainly as drift direction and shape. Please add quantitative error measures on Pohang (and, if possible, on KITTI sequence 06 with known ground truth), including per-sequence numbers and, ideally, error bars or multiple runs. An ablation that isolates the effect of the proposed loss from random fine-tuning would also be needed.","section":"Section III.B, Fig. 5"},{"comment":"The labels for 'good' features are computed from reprojection errors that use the relative pose estimated by RTAB-Map from the same SuperPoint features being trained. This is a potential circular dependency: if the VO pose estimate is biased in the low-texture or dynamic scenes the method targets, features that are consistently wrong under the biased pose can still produce low reprojection error and will be labeled 'good'. The paper provides no evidence that the RTAB-Map poses on Pohang are accurate enough for the labels to be trustworthy. A concrete test is needed: e.g., compare the labels against ground-truth pose on a dataset with known poses, or use an independent pose estimator to generate labels, or demonstrate that the label quality improves as training proceeds. Without this, the training signal may reinforce the original feature extractor's biases rather than correct them.","section":"Section II, Eq. (1)"},{"comment":"The loss is defined only symbolically. The weights w_i, w_i', w_pk, w_d, the reprojection error threshold, and the exact procedure for constructing Y and Y' are not specified. The 'peaky loss' and the descriptor hinge loss L_d are described in one sentence each, which is insufficient to reproduce or analyze the method. Training details (learning rate, number of iterations, dataset split, warm-start, whether both stereo and temporal correspondences are used in each batch) are also omitted. Without these, the reported qualitative results cannot be reproduced or compared against alternative self-supervised schemes.","section":"Section II, Eq. (1)"},{"comment":"The claim that features become 'more equitably spread' and that unreliable regions (terrain, water) are suppressed is made only through qualitative feature maps. There is no quantitative measure of feature distribution (e.g., coverage, cell occupancy, nearest-neighbour spacing), repeatability, track length, or the fraction of features that fall on the water/terrain. Moreover, Fig. 4 is captioned as KITTI seq. 06 while the text in Section III.B says 'see Fig. 4' when discussing Pohang; please clarify which dataset each panel shows and add quantitative distribution statistics for both Pohang and KITTI.","section":"Section III.A, Figs. 3-4"},{"comment":"The authors explicitly state that they did not train the feature matcher SuperGlue, which remains pre-trained on the original SuperPoint, and that 'current results are also not yet sufficient' to answer the paper's central question. Since the title and abstract state that the method improves 'feature extraction and tracking', and the pipeline includes SuperGlue, this is a load-bearing limitation. The claims should either be explicitly restricted to feature extraction, or SuperGlue should be jointly fine-tuned and the trajectory results re-evaluated. As written, the tracking component is not actually modified.","section":"Section IV"}],"minor_comments":[{"comment":"The notation X, X', Y, Y', D, D', S is used without formal definitions. Please define the score map, descriptor map, label map, and the warping operation precisely, including dimensions.","section":"Section II, Eq. (1)"},{"comment":"The legend includes a 'Baseline' trajectory that is never defined in the text. Also, the axis labels and scales are inconsistent across the four subplots, making visual comparison difficult. It would help to plot all methods on the same coordinate scale with a common origin.","section":"Fig. 5"},{"comment":"There are several cross-reference inconsistencies (e.g., 'see Fig. 4' when discussing Fig. 5; Fig. 3 vs Fig. 4 refer to different datasets in different paragraphs). Please correct these and ensure every figure is cited at the point where it is discussed.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is quite preliminary for a journal submission: the only evidence is qualitative, and the authors themselves disclaim sufficiency. The main risk is the circular label-generation loop, which needs a dedicated experiment. I believe the core idea is promising and addressable, so major revision is appropriate rather than rejection. The paper would also benefit from being repositioned as a proof-of-concept for self-supervised feature adaptation, with explicit scope limitations, since the current title overclaims relative to the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a short, honest proof of concept for a plausible idea: fine-tune SuperPoint (and eventually SuperGlue) using reprojection errors computed inside RTAB-Map, so that the feature detector learns to output features that are easy to track in the target environment. The integration is concrete, the loss terms are sensible, and the authors are careful not to overclaim—they repeatedly call it preliminary and say the current results are not sufficient. That restraint earns credit.\n\nWhat is genuinely new is the specific self-supervision signal: the VO's own geometric consistency, rather than an external task or manual labels. That is a legitimate extension of SuperPoint's self-supervised training. The qualitative feature maps do show a visible shift toward rigid structures and away from water/terrain, which suggests the training signal is changing behavior.\n\nThe soft spots are real. There are no quantitative metrics anywhere—no ATE, RPE, or even a table of trajectory errors. The trajectory plots have no ground truth, so 'outperforms' is not supported. Loss weights and training details are omitted. And the stress-test concern is on target: the 'good' labels are computed from the VO's estimated pose, which came from the same features being trained. If that pose is biased in low-texture or dynamic scenes, the reprojection errors can be small for the wrong features, and fine-tuning can amplify the bias. This is a potential invalidation of the training signal in the very scenes the method targets. The authors don't address it.\n\nThat said, the circularity is not necessarily fatal. The labels are geometric consistency checks, not the final trajectory error, and the baseline is the pre-trained SuperPoint, so at worst the method might over-fit to RTAB-Map's failure modes. But the paper needs to demonstrate, with numbers, that the labeled features actually are better—e.g., by evaluating with a different pose solver or comparing against ground truth.\n\nWho is this for? Researchers working on feature learning for VO/SLAM, and people thinking about self-supervised training with downstream task feedback. It would be a good workshop poster and a useful discussion paper. It deserves a serious referee, not a desk reject, because the idea is worthwhile and the authors are transparent. But it needs major revision: add quantitative evaluation, ablations, and a direct examination of the circularity.\n\nFor now, I'd cite it only as related work, not as a validated method.","headline":"A plausible self-supervised fine-tuning idea for feature extractors in VO, honestly presented as a proof of concept, but the evidence is qualitative and the labeling loop is genuinely circular.","tokens_in":5050,"tokens_out":2504,"would_cite":false,"duration_ms":28196,"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":"Self-supervised fine-tuning with a VO's own reprojection errors makes deep features spread more evenly across rigid structures and improves estimated trajectories in a maritime dataset.","keywords":["self-supervised learning","visual odometry","feature extraction","feature tracking","SuperPoint","SuperGlue","maritime robot navigation","RTAB-Map"],"falsifier":"Fix the RTAB-Map pipeline and compare trajectory error on a held-out sequence (e.g., a new Pohang run or KITTI sequence not used for qualitative checks) between the pre-trained and fine-tuned SuperPoint. If the fine-tuned model does not reduce mean trajectory error, or if it concentrates features on rigid structures but worsens pose estimation in textureless offshore scenes, the central claim fails. A quicker probe: add calibrated noise to the relative pose used to label features during training; if the fine-tuned features still spread evenly and still improve trajectories, then reprojection-e","tokens_in":4293,"feed_emoji":"🛥️","tokens_out":6661,"duration_ms":66123,"temperature":0.7,"pith_summary":"The paper argues that deep feature extractors like SuperPoint, though better than handcrafted detectors at spreading features, still misbehave on out-of-distribution outdoor scenes: they waste detections on water, terrain, and other unreliable regions. The authors propose a self-supervised loop in which a visual odometry system (RTAB-Map with SuperPoint and SuperGlue) labels its own reprojection-consistent features as 'good', and these labels fine-tune SuperPoint through a weighted loss on score maps, peakiness, and descriptors. Preliminary results on the Pohang maritime dataset show the fine-tuned model suppresses water and terrain features, spreads detections across rigid structures, and traces qualitatively better vessel trajectories at daytime and nighttime. The contribution is a proof of concept that task-level feedback from a VO can re-shape a generic feature extractor without manually labelled keypoints.","feed_headline":"VO's own reprojection errors retrain features for steadier tracks","feed_subtitle":"Fine-tuning SuperPoint on VO feedback spreads features over rigid structures and improves vessel trajectories in the Pohang dataset.","key_machinery":"The mechanism is a self-supervised training loop over the feature extractor, with a total weighted loss L = w_i·L_i(X,Y) + w'_i·L_i(X',Y') + w_pk·L_pk + w_d·L_d(D,D',S). The labels Y and correspondences come from the VO's own stereo and temporal matching and its estimated relative pose; reprojection errors separate tracked 'good' features from unreliable ones. The peaky loss L_pk penalizes diffuse or ambiguous keypoint score maps (the detector's per-pixel confidence), and the descriptor hinge loss L_d enforces separation of non-matching descriptors. This turns the VO into a label generator for the feature extractor, so the extractor is optimized for the actual geometry of the task rather tha","core_discovery":"The central claim is that 'good deep features to track' can be defined by the downstream task itself: a feature is good if, given the VO's estimated relative pose, it yields low reprojection error and is consistently tracked across stereo and temporal frames. Using this definition, the paper fine-tunes SuperPoint inside a self-supervised pipeline. The updated extractor produces score maps that concentrate on rigid infrastructure and suppress terrain and water surfaces, and the resulting features, fed back into RTAB-Map, give reconstructed vessel trajectories that stay closer to the expected paths in the Pohang dataset than trajectories from pre-trained SuperPoint or ORB features. The paper i","pith_inferences":["The same self-supervision could be iterated: after fine-tuning the extractor, the improved VO poses produce cleaner labels, so multiple rounds might converge to a fixed point where extractor and pose estimator are mutually consistent; a natural extension is to monitor whether such iteration stabilizes or drifts.","Since SuperGlue was left frozen, the loss's descriptor term had no effect on matching; fine-tuning the matcher jointly might amplify the gains, or reveal that descriptor updates are unnecessary when the extractor alone reshapes the score maps.","A quantitative coverage metric (e.g., number of features on water vs. structure, or entropy of feature distribution) could turn the qualitative 'more equitably spread' observation into a testable objective; the authors don't provide one, but the claim implies a measurable redistribution.","The nighttime offshore result—SuperPoint drifts less at night than day—suggests the model implicitly learns to ignore reflections and moving objects; one could probe this by testing on scenes with controlled specularities or dynamic boats."],"forward_implications":["Fine-tuning with task feedback generalizes better than static pre-trained features in environments that look unlike the training data, especially low-texture maritime scenes.","The same label-generation loop can be applied to other VO/VSLAM pipelines and other feature extractors, turning any pose estimator into a supervisor for its own front-end.","Features become more evenly distributed over rigid structures, giving more stable geometric constraints for rotation from distant points and translation from nearby points.","Trajectory improvement appears in both daytime and nighttime narrow-canal sequences; in offshore areas the benefit is less pronounced, suggesting the method helps most when texture is scarce but structures exist.","The proof of concept implies label-intensive feature annotation can be replaced by geometric consistency labels, reducing the cost of adapting detectors to new domains."],"supporting_citations":[{"why":"Supplies the deep feature extractor (SuperPoint) that the paper fine-tunes with self-supervised labels.","marker":"[2]"},{"why":"Supplies the pretrained matcher (SuperGlue) used for temporal and stereo correspondences in the training loop.","marker":"[9]"},{"why":"Provides the RTAB-Map visual odometry pipeline that generates pose estimates and reprojection-error labels and is used for evaluation.","marker":"[11]"},{"why":"Provides the Pohang maritime dataset used for qualitative feature comparison and trajectory evaluation.","marker":"[5]"},{"why":"Serves as the classical ORB-feature baseline trajectory for comparison in all four evaluation scenes.","marker":"[13]"},{"why":"Provides the KITTI sequence used for a qualitative check of feature extraction outside the maritime domain.","marker":"[12]"}],"fun_headline_variants":["Self-supervised VO feedback retrains features for rugged outdoor tracks","VO's own errors fine-tune features for more stable motion tracking","Task-specific self-supervision makes deep features track better in VO","Retrained features from VO feedback improve trajectory accuracy outdoors","Self-supervised fine-tuning of SuperPoint for robust visual odometry"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The training labels for 'good' features are produced by the VO's own estimated relative pose and reprojection errors, so the method assumes those pose estimates are accurate enough in the very scenes—low-texture, dynamic, harsh lighting—where the extractor is supposed to improve.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised VO feedback retrains features for rugged outdoor tracks","VO's own errors fine-tune features for more stable motion tracking","Task-specific self-supervision makes deep features track better in VO","Retrained features from VO feedback improve trajectory accuracy outdoors","Self-supervised fine-tuning of SuperPoint for robust visual odometry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000206,"raw_usage":{"total_tokens":1170,"prompt_tokens":617,"completion_tokens":553,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":361,"completion_tokens_details":{"reasoning_tokens":466}},"tokens_in":361,"tokens_out":553,"duration_ms":5326,"temperature":1.0,"reasoning_tokens":466,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T20:44:30.304111+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix the RTAB-Map pipeline and compare trajectory error on a held-out sequence (e.g., a new Pohang run or KITTI sequence not used for qualitative checks) between the pre-trained and fine-tuned SuperPoint. If the fine-tuned model does not reduce mean trajectory error, or if it concentrates features on rigid structures but worsens pose estimation in textureless offshore scenes, the central claim fails. A quicker probe: add calibrated noise to the relative pose used to label features during training; if the fine-tuned features still spread evenly and still improve trajectories, then reprojection-e","supporting_citations":[],"review_version":1}