{"id":"1b49ea2b-78c5-41c3-944b-f003a83c399c","arxiv_id":"2505.00980","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"LMDepth is a lightweight Mamba-based monocular depth estimation network that claims improved accuracy with fewer parameters and FLOPs on NYUDv2 and KITTI, plus real-time INT8 deployment on embedded hardware.","lead":"This paper describes LMDepth, a compact neural network that estimates depth from a single image using Mamba sequence blocks alongside standard vision components. It reports better accuracy with fewer parameters and FLOPs than several lightweight depth estimators, and demonstrates INT8 deployment on an embedded Jetson platform.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"KITTI comparison in Table II is internally suspect: the FastDepth row reproduces JointDepth's NYUDv2 numbers with columns shifted, so the headline accuracy-efficiency claim is not currently supported by the reported comparisons.","rationale":"The reader's weakest assumption was that baseline metrics are not directly comparable because of unspecified splits and resolutions. I agree with that concern, but the sharper problem is internal: Table II's low-resolution FastDepth row is numerically identical to Table I's JointDepth NYUDv2 row with the final two columns swapped. This is not merely an unstated evaluation protocol; it is positive evidence that the KITTI comparison table is unreliable. That directly threatens the headline claim that LMDepth is best across all metrics on KITTI. At the same time, this is a data-provenance problem, not a demonstrated flaw in the architecture or a contradiction in the proposed method. The SSM formalism is standard, the loss sign in Eq. (5) is likely a typo rather than a conceptual error, and the deployment section's contradictory model sizes (26 MB vs. 8.54 MB in Table V) are secondary to the accuracy claim. The appropriate outcome remains CONDITIONAL: the paper should be accepted only if the authors release code and regenerate all comparisons under a single, clearly specified protocol. Since the reader already reached CONDITIONAL, my stress-test does not move the verdict.","tokens_in":13834,"tokens_out":6636,"duration_ms":69360,"concrete_test":"Audit the provenance of every Table II baseline row. Concretely, recover the original FastDepth and JointDepth KITTI numbers from their papers or codebases and compare them to Table II. If FastDepth's KITTI metrics are not exactly 0.778/0.942/0.987/0.142/0.591/— under the cited protocol, the low-resolution KITTI comparison is corrupted. Then re-run all baselines and LMDepth on the same KITTI Eigen split with the same depth cap and resolution; if corrected δ1 and REL no longer beat DaNet and GuideDepth at 384×1260, the central efficiency-accuracy claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that LMDepth achieves the best performance across all metrics on both NYUDv2 and KITTI rests on the baseline tables, and Table II contains a concrete internal inconsistency. The low-resolution FastDepth row reads δ1=0.778, δ2=0.942, δ3=0.987, REL=0.142, Sq.rel=0.591, RMSE=—, which is numerically identical to Table I's low-resolution JointDepth NYUDv2 row (δ1=0.778, δ2=0.942, δ3=0.987, REL=0.142, —, RMS=0.591) with the last two columns interchanged. This strongly suggests that KITTI baseline metrics were copied or transposed rather than produced under a common evaluation protocol. Because Section IV.A never states the KITTI split (Eigen vs. official), the depth cap, or the preprocessing, and no baseline is re-run, the claimed advantage over DaNet and GuideDepth on KITTI could be an artifact of mismatched or erroneous baseline values. The NYUDv2 comparison may be valid, but the global 'best across all metrics' claim, which explicitly includes KITTI, is not established by the data as printed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LMDepth, a lightweight Mamba-based monocular depth estimation network for embedded deployment. The architecture uses a MobileNetV2 encoder, a modified pyramid spatial pooling (MPSP) module that outputs scene classification and depth bins, and a decoder composed of Depth Mamba Blocks (DMBs). The authors report state-of-the-art accuracy-efficiency trade-offs on NYUDv2 and KITTI, with lower parameters and FLOPs than prior lightweight methods, and additionally present a deployment study on a Jetson AGX Xavier with INT8 quantization and zero-shot generalization on a self-collected indoor dataset. The central claim is that LMDepth achieves the best performance across all metrics among lightweight depth estimation methods on both benchmarks while remaining computationally lighter.","tokens_in":14097,"tokens_out":4223,"duration_ms":45551,"significance":"If the reported results are reliable, LMDepth would be a valuable contribution to efficient monocular depth estimation, extending Mamba-based architectures to dense prediction with a favorable accuracy-cost balance. The paper includes a practical deployment study, which is a strength relative to many architecture papers, and the proposed MPSP and DMB components are clearly described. However, the central comparative claims currently rest on baseline tables that contain at least one concrete internal inconsistency, and the evaluation protocol is not specified in sufficient detail to allow independent verification. The lack of released code or the self-collected evaluation set further limits reproducibility. The architectural idea and empirical direction are promising, but the evidence as printed does not yet fully support the headline 'best across all metrics' claim.","major_comments":[{"comment":"The low-resolution FastDepth row in the KITTI table reports δ1=0.778, δ2=0.942, δ3=0.987, REL=0.142, Sq-rel=0.591, RMSE=—, which is numerically identical to the low-resolution JointDepth row in Table I (NYUDv2) with the last two columns interchanged (JointDepth: δ1=0.778, δ2=0.942, δ3=0.987, REL=0.142, log10=—, RMS=0.591). This indicates that at least one KITTI baseline entry was copied or transposed from a different dataset/table rather than produced under the stated evaluation protocol. Since the paper's central claim of 'best performance across all evaluation metrics' on KITTI depends on the correctness of these baseline numbers, the authors should either correct Table II to match the original FastDepth results on KITTI or re-run the baseline under the same protocol and report the resulting numbers.","section":"Table II, Section IV.B"},{"comment":"No evaluation protocol is specified for either dataset: the train/validation/test split (e.g., the standard Eigen split for KITTI or the official split for NYUDv2), the depth range or cap, the cropping scheme (e.g., Eigen crop, Garg crop, or center crop), and the preprocessing applied to RGB and depth images are all absent. Because no baseline is re-run in the authors' codebase and all baseline numbers are cited from previous papers, differences in these settings can easily account for the reported margins. The paper should state the exact protocol for both datasets and, ideally, verify at least one or two baselines (e.g., GuideDepth and DaNet) under the same protocol before claiming superiority across all metrics.","section":"Section IV.A and Tables I-II"},{"comment":"The architecture variants (projection feature dimension, number of MPSP pooling scales, and the choice of LMDepth vs. LMDepth-S) are selected based on accuracy and FLOPs trade-offs that appear to be evaluated on the same benchmark that is later reported as the main result. The paper does not describe a separate validation split used during model selection. This makes the reported test numbers selection-aware rather than independent, inflating the apparent advantage. The authors should clarify which split was used for the ablation/selection curves and report the corresponding test-set results in a way that does not reuse the selection data.","section":"Section IV.C and Figure 5"},{"comment":"There are contradictory deployment numbers. The text states that converting to ONNX 'reduces the model size from 26MB to 8.54MB', but Table V lists the PyTorch model size as 8.54 MB, with the ONNX size as 4.22 MB and the quantized size as 2.63 MB. Additionally, the text refers to a 'pruned LMDepth-T model' achieving 120 FPS, but no model named LMDepth-T is defined anywhere in the paper, and Table IV lists the same FPS for 'LMDepth (Ours)'. The deployment claims should be corrected and the naming clarified, since the claimed 4x compression and 2x speed-up do not match the numbers in Table V as printed.","section":"Section V.B and Table V"},{"comment":"The classification loss is written as Lcls(y, y_hat) = log(y_hat_k), with y_hat representing the predicted probabilities and y the ground truth. As a loss to be minimized, this should be the negative log-likelihood, i.e., -log(y_hat_k) for the true class k. If the implementation actually uses the negative log (as is standard for cross-entropy), the equation should be corrected; as written, minimizing Lcls would drive the predicted probability of the true class toward zero, which is inconsistent with the reported successful training. This is a load-bearing reproducibility issue.","section":"Equation (5), Section III.E"}],"minor_comments":[{"comment":"The text says 'far lower than methods like GuideDepth (5.72 GFLOPs)', but Table I lists the high-resolution GuideDepth FLOPs as 10.47 G (and low-resolution as 2.63). Please correct this number to match the table.","section":"Section IV.B, paragraph after Table I"},{"comment":"The heading 'Data Acquisiotion Device' contains a typo; it should be 'Data Acquisition Device'.","section":"Section V.A"},{"comment":"The term 'pruned LMDepth-T' is introduced without any prior definition or explanation of pruning; if a pruned model is used for the FPS comparison, its construction should be described, or the name should be changed to match the model in Table IV.","section":"Section V.B"},{"comment":"The figure caption and text use inconsistent capitalization of 'MAMBA' (all caps) versus 'Mamba'; standardize to the conventional 'Mamba'.","section":"Figure 2, Section I"},{"comment":"The dataset descriptions are very brief: the KITTI description does not mention the number of scenes or frames used, and the NYUDv2 description does not specify the split (e.g., 249 training scenes / 654 test images commonly used). Adding these details would improve reproducibility.","section":"Section IV.A, datasets"}],"recommendation":"major_revision","confidential_remarks":"The internal inconsistency in Table II (FastDepth row matching another table's JointDepth row with swapped columns) is a serious red flag for the reliability of the baseline comparison. I recommend that the editor request the authors provide the original evaluation logs or re-run the key baselines under a common protocol before this paper is considered further. The missing evaluation protocol and model selection without a held-out validation set are also standard empirical-validation concerns that, while fixable, currently prevent the headline claims from being trusted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick note on LMDepth. The headline claim—best accuracy-efficiency point among lightweight MDE methods on NYUDv2 and KITTI—is not supported by the tables as printed. The stress-test note is right: the low-res FastDepth row in Table II (δ1=0.778, δ2=0.942, δ3=0.987, REL=0.142, Sq-rel=0.591) is identical to the JointDepth row in Table I, with the last two columns interchanged. That is not a coincidence; it looks like KITTI baseline numbers were copied from the NYUDv2 table. Since Section IV.A never states the KITTI split, depth cap, or preprocessing, the claimed margins over DaNet and GuideDepth on KITTI are unverifiable and likely artifacts. The NYUDv2 comparison may be valid, but the global claim has to be walked back.\n\nThat said, the core idea is not bad. The architecture is a sensible combination: MobileNetV2 encoder, PSP-style pooling with adaptive bins, and VMamba decoder blocks. The ablation in Table III—Mamba vs. Transformer vs. CNN decoder—shows Mamba getting slightly better δ1 than Transformer with a fraction of the FLOPs, which is exactly the kind of evidence the community needs. If that holds up, it is a useful data point for edge depth estimation.\n\nSoft spots beyond the KITTI table: no error bars anywhere, no validation split described, and the deployment prose contradicts Table V (the text says 26MB→8.54MB and 'over 4× compression,' but Table V lists 8.54MB as the starting point and 2.63MB as quantized, which is roughly 3.2×). The 'first lightweight Mamba-based MDE' claim is unsupported because no prior Mamba depth work is cited, and the self-collected evaluation set is not released.\n\nMy take: this needs heavy revision before it can be trusted. The architecture and ablation deserve referee time, but the current comparison is not honest evidence. I would send it to review with a strong request to re-run baselines or clearly state the protocol, and to fix the deployment numbers. If the authors can produce code and corrected tables, there is a decent paper here. As is, cite with caution.","headline":"The KITTI baseline table appears to contain NYUDv2 numbers copied with columns swapped, so the headline claim is currently unsupported, though the architecture and ablation are worth a second look.","tokens_in":14609,"tokens_out":3811,"would_cite":false,"duration_ms":34719,"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":"LMDepth, a lightweight Mamba-based monocular depth estimator, claims the best accuracy among lightweight methods on NYUDv2 and KITTI while using fewer parameters and lower FLOPs.","keywords":["monocular depth estimation","Mamba","state space models","lightweight network","depth bins","INT8 quantization","embedded deployment"],"falsifier":"Re-run the compared lightweight baselines (FastDepth, GuideDepth, DaNet, TuMDE) on the same train/validation split and resolutions used for LMDepth on NYUDv2 and KITTI without changing anything else; if any baseline matches or exceeds LMDepth's δ1 or REL at equal or lower FLOPs, the central claim of a new efficiency frontier is falsified. A complementary check is to measure actual per-model latency on the same embedded board, since GFLOPs do not always translate to speed.","tokens_in":13630,"feed_emoji":"📷","tokens_out":6497,"duration_ms":63070,"temperature":0.7,"pith_summary":"This paper tries to establish that a state-space (Mamba) decoder can replace the heavy attention modules of Transformer-based depth estimators without sacrificing accuracy. Its model, LMDepth, combines a MobileNetV2 encoder with a modified pyramid spatial pooling head that outputs scene class, depth bins, and global features, and a decoder built from Depth Mamba Blocks. On NYUDv2 and KITTI, the authors report the best metrics among lightweight methods while using fewer parameters and GFLOPs (for example, δ1 = 0.854 and REL = 0.123 at 480×640 with 2.9M parameters and 2.77 GFLOPs). They further show that INT8 quantization shrinks the model to 2.63 MB with minimal accuracy loss. If correct, this gives resource-constrained robots and autonomous systems a depth-estimation backbone that is both accurate and cheap to run.","feed_headline":"Mamba depth net beats lightweight baselines at lower FLOPs","feed_subtitle":"On NYUDv2 and KITTI, LMDepth scores top accuracy with under 3M parameters, then runs INT8-quantized on an embedded GPU.","key_machinery":"The Depth Mamba Block (DMB) is the carrier of the argument: a decoder unit built on Vmamba modules that process image and depth tokens with a one-dimensional convolution followed by a selective state-space model, giving a global receptive field with cost linear in sequence length. Alongside it, the Modified Pyramid Spatial Pooling (MPSP) head pools at scales {1, 2, 3, 6} (or {1, 6} in the small variant) to produce scene classification, depth distribution bins, and global features. The mechanism's work is to let a lightweight network capture global context and discretize depth into bins, then reconstruct depth as probability-map-weighted bin values.","core_discovery":"The central claim is that LMDepth achieves the best performance across all evaluation metrics among lightweight monocular depth estimation methods on both NYUDv2 and KITTI, while requiring fewer parameters and lower GFLOPs. The architecture gets this from two components: a Modified Pyramid Spatial Pooling (MPSP) head that aggregates multi-scale context and predicts scene classification plus depth bins, and a decoder of Depth Mamba Blocks (DMBs) that fuses depth and visual features through linear-complexity state-space layers. The final depth is obtained by element-wise multiplication of a per-pixel probability map with the predicted bins. The authors also report that ONNX conversion plus INT8 quantization compresses the model to 2.63 MB and raises inference speed to 122 FPS on an embedded GPU platform, supporting deployment on resource-constrained devices.","pith_inferences":["Editorial inference: Because the MPSP head outputs scene classification, the same backbone could naturally extend to joint depth estimation and scene understanding; the paper only uses classification as an auxiliary loss.","Editorial inference: The reported efficiency gains are measured in parameters and GFLOPs; actual latency on other edge devices, especially CPUs or lower-power NPUs, may not scale identically because Mamba's hardware-aware efficiency is partly GPU-oriented.","Editorial inference: The bin-based depth representation suggests a direct route to uncertainty estimation: the probability map over bins could be reinterpreted as a per-pixel depth distribution, yielding confidence intervals without extra supervision.","Editorial inference: A head-to-head re-run of baselines under identical splits, resolutions, and preprocessing would settle whether the reported margin is architectural or partly due to protocol differences; the paper relies on published numbers."],"forward_implications":["Mamba-style linear state-space decoders can replace quadratic-attention decoders in dense prediction: on NYUDv2, the Transformer decoder needs 2.8 GFLOPs versus 0.73 GFLOPs for the Mamba decoder, with similar or better accuracy.","A lightweight depth network can keep global context: the MPSP and DMB design reaches δ1 = 0.854 on NYUDv2 at 480×640 with only 2.9M parameters, a point on the accuracy-efficiency frontier.","The same network transfers across indoor and outdoor benchmarks: NYUDv2 δ1 = 0.854 and KITTI δ1 = 0.926 at high resolution, and zero-shot generalization on self-collected scenes beats the compared lightweight methods.","Practical edge deployment is feasible: ONNX plus INT8 quantization reduces the model to about 10% of its original size and reaches 122 FPS on an embedded platform, with δ2 dropping only from 0.86 to 0.84.","The auxiliary scene-classification branch provides a lightweight way to guide depth-bin generation, suggesting that semantic context and depth prediction can share a single compact head."],"supporting_citations":[{"why":"Supplies the selective state space model that gives Mamba its linear-time sequence modeling, the foundation of the Depth Mamba Block.","marker":"[29]"},{"why":"Provides the Vmamba module that each Depth Mamba Block integrates to fuse depth and visual features.","marker":"[16]"},{"why":"MobileNetV2 serves as the image encoder backbone that extracts the multi-scale features fed into MPSP and the decoder.","marker":"[33]"},{"why":"AdaBins supplies the adaptive depth-bin formulation and the loss settings (λ = 0.85, α = 10) that LMDepth adopts.","marker":"[9]"},{"why":"Defines the scale-invariant loss and the six evaluation metrics used for all main comparisons.","marker":"[21]"},{"why":"FastDepth is a classic lightweight CNN depth estimator used as a baseline for accuracy, FLOPs, and FPS comparisons.","marker":"[12]"},{"why":"GuideDepth is the strongest lightweight CNN baseline on NYUDv2 and KITTI that LMDepth claims to beat at high resolution.","marker":"[14]"},{"why":"DaNet is the transformer-based lightweight baseline used to show that Mamba achieves comparable or better accuracy at far lower cost.","marker":"[37]"},{"why":"NYUDv2 is the indoor dataset and provides the official 25-class labels used for the auxiliary scene-classification task.","marker":"[40]"},{"why":"KITTI is the outdoor dataset used to demonstrate that the efficiency and accuracy gains hold outside indoor scenes.","marker":"[41]"}],"fun_headline_variants":["Mamba depth net: beats lightweight baselines, fewer FLOPs","LMDepth: linear-complexity Mamba depth, top accuracy on KITTI","Lightweight Mamba depth: high accuracy, low FLOPs, edge-ready","LMDepth: 2.63 MB, 122 FPS on embedded GPU","Mamba depth estimation: fewer params, better depth, edge deployable"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported advantage rests on the assumption that the baseline numbers cited from earlier papers were produced under protocols equivalent enough to LMDepth's — same data split, resolution, and preprocessing — so that cross-table comparisons are apples-to-apples.","fun_headline_variants_meta":{"raw":{"variants":["Mamba depth net: beats lightweight baselines, fewer FLOPs","LMDepth: linear-complexity Mamba depth, top accuracy on KITTI","Lightweight Mamba depth: high accuracy, low FLOPs, edge-ready","LMDepth: 2.63 MB, 122 FPS on embedded GPU","Mamba depth estimation: fewer params, better depth, edge deployable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00076,"raw_usage":{"total_tokens":3375,"prompt_tokens":946,"completion_tokens":2429,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":2326}},"tokens_in":562,"tokens_out":2429,"duration_ms":17785,"temperature":1.0,"reasoning_tokens":2326,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:29:50.275975+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the compared lightweight baselines (FastDepth, GuideDepth, DaNet, TuMDE) on the same train/validation split and resolutions used for LMDepth on NYUDv2 and KITTI without changing anything else; if any baseline matches or exceeds LMDepth's δ1 or REL at equal or lower FLOPs, the central claim of a new efficiency frontier is falsified. A complementary check is to measure actual per-model latency on the same embedded board, since GFLOPs do not always translate to speed.","supporting_citations":[{"cited_title":"Adabins: Depth estimation using adaptive bins,","cited_arxiv_id":null,"evidence_quote":"AdaBins supplies the adaptive depth-bin formulation and the loss settings (λ = 0.85, α = 10) that LMDepth adopts."},{"cited_title":"Fastdepth: Fast monocular depth estimation on embedded systems,","cited_arxiv_id":null,"evidence_quote":"FastDepth is a classic lightweight CNN depth estimator used as a baseline for accuracy, FLOPs, and FPS comparisons."},{"cited_title":"Lightweight monocular depth estimation through guided decoding,","cited_arxiv_id":null,"evidence_quote":"GuideDepth is the strongest lightweight CNN baseline on NYUDv2 and KITTI that LMDepth claims to beat at high resolution."},{"cited_title":"Monocular depth distribution alignment with low computation,","cited_arxiv_id":null,"evidence_quote":"DaNet is the transformer-based lightweight baseline used to show that Mamba achieves comparable or better accuracy at far lower cost."}],"review_version":1}