{"id":"55727eeb-2e67-438f-b106-1690363635c9","arxiv_id":"2411.13988","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"DU-VIO combines a GAN dehazing module with a CNN-LSTM visual-inertial odometry pipeline and reports improved pose RMSE on modified AQUALOC underwater sequences.","lead":"The paper builds a visual-inertial pose estimator for underwater robots by adding a GAN-based dehazing step before a hybrid CNN-LSTM network. It tests the system on degraded versions of a small underwater dataset and reports lower pose error with dehazing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Dehazing GAN training split is not shown to exclude test sequences h01/h07; if it included them, DU-VIO's gain over U-VIO is explained by data leakage, not by dehazing.","rationale":"The reader's weakest assumption identifies the same load-bearing concern I find: the dehazing GAN's training set is not documented as excluding the pose-estimation test sequences. The paper's strongest claim is precisely that dehazing preprocessing improves pose accuracy, so any contamination of the dehazing module with test data directly undermines that claim. Section IV-C's '80% whole images from the modified AQUALOC dataset' strongly suggests a random global split across all seven sequences, and Section IV-B defines h01 and h07 as the test sequences for the pose model. The paper provides no split file, no hash-based overlap check, and no statement that h01/h07 were held out from dehazing training. This is not merely a missing detail; it is a confound that can fully explain the observed improvement. The concrete test I propose would settle the question: train dehazing only on the pose-training/validation sequences and see whether DU-VIO still beats U-VIO on h01/h07. If the advantage vanishes, the central claim is an artifact; if it persists, the paper's conclusion would be supported. I agree with the reader's verdict of REJECT because the evaluation, as written, does not establish the claim. A conditional acceptance would require the re-evaluation described above, which is why I keep the verdict unchanged rather than moving it.","tokens_in":9651,"tokens_out":2611,"duration_ms":25156,"concrete_test":"Retrain the dehazing GAN using only images from sequences h02, h03, h04, h05, and h06 (or any split that provably excludes h01 and h07), then re-evaluate DU-VIO versus U-VIO on h01 and h07 under all three scenarios. If the translation or rotation RMSE improvement shrinks or disappears, the original DU-VIO gain is a training-selection artifact; if the improvement persists, the dehazing benefit is confirmed. Also report the dehazing training split and verify non-overlap via image hashes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that adding the GAN dehazing module improves pose estimation over the same VIO backbone without it (DU-VIO vs U-VIO). This claim rests on the test sequences h01 and h07, as defined in Section IV-B. Section IV-C states that the dehazing module was trained on '80% whole images from the modified AQUALOC dataset' with no split specification. The modified AQUALOC dataset comprises the seven harbor sequences, so a random 80% selection almost certainly includes frames from h01 and h07. If those test frames were seen by the dehazing GAN during training, the dehazed test images are no longer representative of unseen underwater scenes; the reported RMSE reductions in Figures 5-8 could reflect the dehazing network's memorization of the test sequences rather than a generalizable visibility-enhancement benefit. The pose estimation module itself is trained only on h02, h04, h06, so the test set is clean for that component; the suspected leakage is specific to the dehazing preprocessing step, which is exactly the variable whose effect the comparison isolates. Without evidence that h01 and h07 were excluded from the dehazing training corpus, the paper's main contribution is not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DU-VIO, a dehazing-aided visual-inertial odometry framework for underwater pose estimation. DU-VIO preprocesses raw camera images with a GAN-based dehazing module and then feeds the enhanced images, together with raw IMU data, into a hybrid CNN-LSTM pose estimation network derived from VS-VIO. The central claim is that adding the dehazing module reduces translation and rotation RMSE on two test sequences (h01 and h07) from the AQUALOC harbor dataset under original, distortion, and turbidity scenarios, comparing DU-VIO with a U-VIO baseline that omits dehazing. The paper also reports an ablation of dehazing backbones, a comparison with other dehazing methods, a comparison with geometric and data-driven VIO baselines, and hardware metrics for inference.","tokens_in":9900,"tokens_out":4157,"duration_ms":39883,"significance":"If the reported comparison were clean, the paper would provide a useful empirical data point for the underwater robotics community: an end-to-end learning-based VIO system augmented by a visibility-enhancement preprocessing module, evaluated on a public underwater dataset. The paper gives credit where due by building explicitly on VS-VIO [9] and its own dehazing module [23], and it includes a backbone ablation, a dehazing state-of-the-art comparison, and hardware resource measurements that are not always reported in VIO papers. However, the central empirical claim is currently compromised by an unresolved overlap between the dehazing training corpus and the pose-estimation test sequences, and the evaluation is otherwise thin (two test sequences, no error bars, interpolated ground truth). The work in its present form does not convincingly establish that dehazing improves pose estimation in extreme underwater conditions.","major_comments":[{"comment":"The dehazing module is trained on \"80% whole images from the modified AQUALOC dataset\" (Section IV-C), while h01 and h07 are designated as the pose-model test sequences in Section IV-B. No information is given about whether the dehazing training split excluded h01 and h07. Since the modified dataset consists of the seven harbor sequences, a random 80% split almost certainly includes frames from h01 and h07. This matters because the paper's central comparison is DU-VIO versus U-VIO, i.e., the same pose estimator with and without the dehazing preprocessor; if the dehazing GAN saw test-sequence frames during training, the RMSE reductions reported in Figures 5-8 may reflect memorization rather than a generalizable dehazing benefit. The authors must report the dehazing data split and either verify that h01 and h07 were excluded or retrain the dehazing module on a split that excludes them and rerun the pose evaluation.","section":"Section IV-C and Section IV-B"},{"comment":"The evaluation rests on only two test sequences, h01 and h07, with each sequence divided into three sub-sequences, and no error bars, confidence intervals, or repeated-run statistics are provided. Section IV-A states that ground truth poses are linearly interpolated from Colmap poses computed on a subset of images (one in five), and only one-third of each trajectory is used for pose estimation. With this small and smoothed test set, the U-VIO versus DU-VIO differences in Figures 5-8 could lie within run-to-run or interpolation-induced variability. The authors should provide aggregate statistics across multiple training runs and, ideally, additional test sequences before claiming that the dehazing module consistently improves pose accuracy.","section":"Section IV-A and Section V"},{"comment":"The state-of-the-art comparison in Table III is reported only for the original scenario, only for DU-VIO, and not for the U-VIO baseline. The table caption also does not state whether the reported numbers are translation RMSE, rotation RMSE, or a combined error, nor whether they are normalized by trajectory length. Without this information, and without results for the distortion and turbidity scenarios, the superiority claim in the text is not fully supported. At minimum, the same comparison should be provided for all three scenarios and for the U-VIO baseline.","section":"Table III"}],"minor_comments":[{"comment":"There is a typo: \"foe 50 epochs\" should be \"for 50 epochs,\" and \"80% whole images\" should be \"80% of the whole images.\"","section":"Section IV-C"},{"comment":"The phrase \"improve textile and content representations\" should likely be \"improve texture and content representations.\"","section":"Section II-A"},{"comment":"The Adam optimizer parameters are written as \"alpha = 0.9 and beta = 0.999\"; in Adam these are conventionally beta1 and beta2, while alpha usually denotes the learning rate, which was already introduced as 1e-6. Please clarify the notation.","section":"Section IV-B"},{"comment":"The sentence \"The Adam optimizer ... is used due to due to its comparatively lower memory demand\" contains a duplicated \"due to.\"","section":"Section IV-B"},{"comment":"The axes in Figures 5-8 are not labeled with units, and the sub-sequence indexing is not defined in the text; please add units and explain how the three sub-sequences are formed.","section":"Section V"},{"comment":"The text uses both \"harbor site\" (Section IV) and \"harbour site\" (Section V); please standardize the spelling.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The central idea is plausible and the authors are transparent about their use of VS-VIO [9] and their own dehazing module [23], but I would require a re-run of the dehazing training with a clean test-excluded split and a more complete statistical evaluation before this paper is publishable. The small test set and the interpolated ground truth are secondary but should also be addressed in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about arXiv:2411.13988. First, it builds a plausible pipeline: dehaze images with a GAN, then feed them to a CNN-LSTM VIO. Second, the evaluation does not support the headline claim because the dehazing GAN appears to have been trained on 80% of the full modified dataset, with no indication that the h01 and h07 test sequences were excluded. If they weren't, the reported DU-VIO gain over U-VIO could be memorization, not generalization.\n\nWhat's actually new: combining a dehazing preprocessor with a learning-based VIO is not, to my knowledge, previously reported for underwater pose estimation. The authors also document hardware metrics (inference time, power, GPU usage) and run a backbone ablation for the dehazing module, which is more than many papers in this area do. They are transparent about using interpolated ground truth and one-third of each trajectory, and they cite the prior work they build on.\n\nThe soft spots are significant. The data-leak issue is load-bearing because the DU-VIO vs U-VIO comparison isolates exactly the dehazing module; if that module saw the test frames, the measured improvement is not valid evidence for dehazing's benefit. The test set is tiny (two sequences), ground truth is interpolated from Colmap, there are no error bars, and the Table III comparison to OKVIS and ORB-SLAM3 is underspecified: did those baselines see the same dehazed images, same trajectories, same training data? The paper does not say. Novelty is also modest: it replaces VS-VIO's policy network with an existing GAN from the authors' own prior work. That is a legitimate extension, not a new method.\n\nWho is this for: people working on learning-based VIO in degraded underwater conditions, and anyone who wants a cautionary example of training/test leakage in a supposedly controlled comparison. I would send it to reviewers, because the core question is worth asking and the flaw is fixable with a clean re-split of the dehazing training data. But as written, the central claim is not well supported.","headline":"A sensible dehazing-aided VIO idea, but the main comparison is likely contaminated because the dehazing GAN appears to have trained on the test sequences.","tokens_in":10492,"tokens_out":2264,"would_cite":false,"duration_ms":21451,"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":"Adding a GAN-based dehazing stage to a learning-based VIO pipeline reduces pose error in extreme underwater conditions on the AQUALOC dataset.","keywords":["Underwater Image Enhancement","Visual-Inertial Odometry","Pose Estimation","Multi-Modal Multi-Rate Data Fusion","Hybrid CNN-LSTM","Dehazing","AQUALOC Dataset","Underwater Robotics"],"falsifier":"Retrain the dehazing module using only the h02–h06 training/validation split, then rerun DU-VIO on h01 and h07 across all three scenarios; if the translation and rotation RMSE improvement over U-VIO disappears or drops sharply, the claimed benefit is an artifact of test-set leakage rather than general dehazing.","tokens_in":9403,"feed_emoji":"🌊","tokens_out":6852,"duration_ms":55903,"temperature":0.7,"pith_summary":"This paper introduces DU-VIO, a visual-inertial odometry framework that prepends a GAN-based dehazing module to a hybrid CNN-LSTM pose estimator, and argues that this front-end improves pose accuracy in extreme underwater scenes with turbidity, distortion, and low texture. The framework takes raw camera images, removes visual disturbances with a DenseNet-121 based generator, then fuses visual features from FlowNet with IMU features in an LSTM that regresses 6-DoF pose. On the AQUALOC harbor sequences h01 and h07, the authors report that DU-VIO achieves lower translation and rotation RMSE than the same framework without dehazing (U-VIO) across original, distorted, and turbid scenarios. They also report that DU-VIO outperforms geometry-based baselines (OKVIS, ORB-SLAM3) and a data-driven baseline (VINet) on the original scenario. If correct, the work suggests that learned dehazing can be a practical front-end for underwater VIO, improving localization for robots in murky water.","feed_headline":"Dehazing before VIO reduces underwater pose error","feed_subtitle":"A learned dehazing module lowers translation and rotation RMSE on AQUALOC test sequences versus the same VIO without dehazing.","key_machinery":"The central object is DU-VIO, a multi-rate multi-modal VIO pipeline whose distinguishing component is a GAN-based dehazing front-end that cleans raw underwater frames before they reach the pose network. The generator uses a DenseNet-121 encoder-decoder with skip connections, and the discriminator judges whether a generated image looks clear. The pose estimator is a hybrid CNN-LSTM: FlowNetSimple extracts visual features from two consecutive dehazed frames, three 1D convolutional layers extract inertial features from the 11 IMU readings between those frames, the two feature vectors are concatenated, and a two-layer LSTM followed by a two-layer MLP regresses the 6-DoF pose at each time step. Multi-rate fusion is built into the design by using a fixed window of high-rate IMU samples per low-rate image pair.","core_discovery":"The central claim is that adding a learned dehazing stage before a visual-inertial odometry network materially improves pose estimation in visually degraded underwater environments. DU-VIO builds on VS-VIO by replacing its policy network with a GAN-based visibility enhancement module, so the visual encoder always receives dehazed, higher-contrast frames. The paper evaluates this by comparing DU-VIO against U-VIO (the same network without dehazing) on the modified AQUALOC dataset under three scenarios—original low-texture, added distortion, and added turbidity—and reports that the dehazed version yields lower translation and rotation RMSE on the two test sequences in most sub-sequence comparisons. It further reports that DU-VIO's RMSE on the original scenario is lower than the compared baselines OKVIS, ORB-SLAM3, and VINet. The paper also documents inference speed and GPU resource usage, positioning DU-VIO as a deployable pipeline.","pith_inferences":["A fair test of the dehazing benefit would retrain the GAN on only the h02–h06 split and test on h01/h07; the paper does not state that h01/h07 images were excluded from the dehazing training set, so the reported improvement may partly reflect exposure to test scenes.","If the dehazing front-end generalizes across sites, the same preprocessing could benefit other downstream underwater tasks such as SLAM, object detection, or segmentation, since all depend on visible features in turbid water.","The multi-rate fusion design (11 IMU samples per image pair) is a reusable recipe for other learning-based VIO systems where sensor rates differ by an order of magnitude.","The authors' removal of the policy network from VS-VIO suggests that in static or slow underwater motion, selective modality gating may be unnecessary; testing this claim would require comparing DU-VIO against VS-VIO with the policy network retained."],"forward_implications":["If DU-VIO's results hold, adding a learned dehazing front-end can improve existing learning-based VIO systems for underwater vehicles without redesigning the pose network.","The framework's performance on distortion and turbidity scenarios suggests the dehazing step helps recover features the visual encoder would otherwise lose, which could extend to other degraded visual conditions such as fog or murky water.","The reported hardware metrics (about 40 seconds inference time, 47 W power, 4% GPU utilization on an A100) indicate the pipeline is light enough for onboard deployment, though real-time operation depends on sequence length and hardware.","DU-VIO's lower RMSE than OKVIS, ORB-SLAM3, and VINet on the original scenario implies that data-driven multi-modal VIO with a learned enhancement module can compete with classical geometry-based methods in underwater settings."],"supporting_citations":[{"why":"Supplies the base VS-VIO CNN-LSTM architecture and training strategy that DU-VIO modifies by dropping the policy network and adding dehazing.","marker":"[9]"},{"why":"Provides the FlowNetSimple convolutional architecture used as the visual feature encoder for two consecutive frames.","marker":"[21]"},{"why":"Supplies the GAN dehazing module design and parameter selection that DU-VIO adopts for visibility enhancement.","marker":"[23]"},{"why":"Provides the AQUALOC dataset, including the harbor sequences, IMU data, and ground-truth poses used for training and evaluation.","marker":"[24]"},{"why":"VINet is the data-driven VIO baseline that DU-VIO is compared against in Table III.","marker":"[5]"},{"why":"OKVIS is the geometry-based VIO baseline that DU-VIO is compared against in Table III.","marker":"[17]"},{"why":"ORB-SLAM3 is the geometry-based VIO baseline that DU-VIO is compared against in Table III.","marker":"[35]"}],"fun_headline_variants":["Dehazing boosts underwater VIO accuracy","Learned dehazing cuts pose error for submersibles","GAN dehazing improves visual-inertial odometry","Underwater VIO gets clearer with dehazing step","Dehazing module sharpens underwater pose tracking"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported benefit of dehazing assumes the dehazing network was not trained on the same h01 and h07 images used to test pose accuracy; Section IV-C says the GAN was trained on 80% of the entire modified dataset, and the paper does not show these test sequences were excluded.","fun_headline_variants_meta":{"raw":{"variants":["Dehazing boosts underwater VIO accuracy","Learned dehazing cuts pose error for submersibles","GAN dehazing improves visual-inertial odometry","Underwater VIO gets clearer with dehazing step","Dehazing module sharpens underwater pose tracking"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000248,"raw_usage":{"total_tokens":1564,"prompt_tokens":982,"completion_tokens":582,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":508}},"tokens_in":598,"tokens_out":582,"duration_ms":4992,"temperature":1.0,"reasoning_tokens":508,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:39:35.056062+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the dehazing module using only the h02–h06 training/validation split, then rerun DU-VIO on h01 and h07 across all three scenarios; if the translation and rotation RMSE improvement over U-VIO disappears or drops sharply, the claimed benefit is an artifact of test-set leakage rather than general dehazing.","supporting_citations":[{"cited_title":"Efficient deep visual and iner- tial odometry with adaptive visual modality selection,","cited_arxiv_id":null,"evidence_quote":"Supplies the base VS-VIO CNN-LSTM architecture and training strategy that DU-VIO modifies by dropping the policy network and adding dehazing."},{"cited_title":"Flownet: Learning optical flow with convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Provides the FlowNetSimple convolutional architecture used as the visual feature encoder for two consecutive frames."},{"cited_title":"Vision-Based Autonomous Navigation for Unmanned Surface Vessel in Extreme Marine Conditions","cited_arxiv_id":"2308.04283","evidence_quote":"Supplies the GAN dehazing module design and parameter selection that DU-VIO adopts for visibility enhancement."},{"cited_title":"Vinet: Visual-inertial odometry as a sequence-to-sequence learning problem,","cited_arxiv_id":null,"evidence_quote":"VINet is the data-driven VIO baseline that DU-VIO is compared against in Table III."},{"cited_title":"Keyframe-based visual–inertial odometry using nonlinear optimization,","cited_arxiv_id":null,"evidence_quote":"OKVIS is the geometry-based VIO baseline that DU-VIO is compared against in Table III."}],"review_version":1}