{"id":"2aa91dc2-41a2-4cf8-a02f-194203a3a0f5","arxiv_id":"2412.20911","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A LiDAR-to-camera distillation method that supervises relative depth inside object foregrounds and distills BEV feature relationships to boost camera-only 3D object detection.","lead":"This paper proposes a training method that transfers geometric knowledge from LiDAR-based teacher detectors into camera-only 3D object detectors for autonomous driving. The method improves camera-based detection accuracy on the nuScenes benchmark, reporting a state-of-the-art 62.8% NDS.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Core distillation equations are dimensionally inconsistent: Eq. (5) and Eq. (7) define losses that cannot produce the stated C×C and N×N relationship matrices given f ∈ R^{N×C}; this undermines the method specification as written.","rationale":"The reader's verdict was already CONDITIONAL, with the weakest assumption identified as unreliable foreground supervision under occlusion. My stress-test pass found a different, more concrete load-bearing concern: the inner-feature BEV distillation, which is one of the two core contributions, is specified with dimensionally inconsistent equations. Given f ∈ R^{N×C}, the formula in Eq. (5) yields an N×N matrix while the text claims C×C, and Eq. (7) yields C×C while the text claims N×N. This is not a subtle implementation detail; it is the formal definition of the proposed loss. The central claim of the paper is that this distillation scheme significantly boosts camera-only detectors, and the ablations in Table V attribute specific gains to LIC_bev and LIK_bev. If those formulas are swapped or mis-stated, the paper's description of its own method is incorrect and the reported ablation cannot be interpreted. The empirical results may still be valid, and the availability of code means the ambiguity is resolvable, but as written the paper does not meet the standard of a reproducible method. I therefore agree with the reader's CONDITIONAL recommendation, but for a more fundamental reason: the method definition itself needs correction or code verification before the central claim can be fully assessed. This is not an attack on the authors; it is a precise technical inconsistency that a concrete code inspection can settle. If the code matches the equations literally, then the loss names in the paper are wrong; if the code matches the text's dimensional claims, then the equations are wrong. Either way, the manuscript must be revised. The occlusion concern raised by the reader is real but secondary; the paper even acknowledges it as a limitation, and the small-object and distance-range ablations suggest the method remains useful in those regimes. The dimensional inconsistency, by contrast, is unacknowledged and sits at the heart of the proposed inner-feature module.","tokens_in":20116,"tokens_out":4813,"duration_ms":51214,"concrete_test":"Inspect the released code at https://github.com/Public-BOTs/TiGDistill-BEV.git and check the actual shapes and operations used for the two distillation losses: (a) confirm whether f_j is N×C or C×N; (b) for the loss called inter-channel, determine whether the code computes f @ f.T or f.T @ f; (c) do the same for the loss called inter-keypoint. If the code computes f.T @ f for inter-channel and f @ f.T for inter-keypoint, then Eq. (5) and Eq. (7) are reversed in the paper and must be corrected. Additionally, rerun the Table V ablation with the two loss terms exchanged: if the +2.1 NDS gain currently attributed to inter-keypoint appears when only the channel-correlation loss is used, the attribution in the paper is inverted.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In Section III-C1, the keypoint features are defined as f_cam_j, f_teacher_j ∈ R^{N×C}, where N is the number of keypoints and C is the number of channels. Equation (5) then defines A_j = f_j f_j^T and claims A_j ∈ R^{C×C} for 'inter-channel' distillation. But for an N×C matrix, f_j f_j^T is N×N, not C×C; the channel-correlation matrix would be f_j^T f_j. Conversely, Equation (7) defines B_j = f_j^T f_j and claims B_j ∈ R^{N×N} for 'inter-keypoint' distillation, but f_j^T f_j is C×C, not N×N. Thus the two formulas are swapped relative to both their names and their stated dimensions. If the implementation follows the equations literally, then the loss labeled LIC_bev actually compares keypoint-correlation matrices and LIK_bev compares channel-correlation matrices, reversing the paper's attribution of which loss captures which geometric relation. If the implementation instead follows the dimensional claims, then Eq. (5) and Eq. (7) are written incorrectly and the paper does not specify the method. Either way, a reader cannot reproduce the inner-feature distillation module from the manuscript, and the ablation in Table V cannot be mapped to the described components. This is an internal inconsistency in a core contribution, not a disagreement with external consensus, and it directly affects the central claim that the proposed losses produce the reported gains.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TiGDistill-BEV, a training-time knowledge-distillation framework for camera-based multi-view BEV 3D object detection. The method adds two components to a camera-only student detector: (1) an inner-depth supervision module that computes relative depth within foreground targets using an adaptively selected reference point, and (2) an inner-feature BEV distillation module that samples keypoints in ground-truth boxes and distills inter-channel and inter-keypoint similarity matrices from a LiDAR or fusion teacher. Experiments on nuScenes report consistent gains over BEVDet, BEVDet4D, and BEVDepth, with a test-set result of 62.8% NDS and 53.9% mAP when using a BEVFusion teacher.","tokens_in":20422,"tokens_out":15009,"duration_ms":146865,"significance":"If the empirical results hold, the paper would demonstrate a practical and reusable recipe for transferring LiDAR or fusion teacher knowledge to camera-only BEV detectors, with gains that are consistent across backbones, temporal settings, and teacher models. The paper's strengths include evaluation on an external benchmark, a large set of ablation tables (Tables III-XIII), multiple baseline detectors, and a public code link. The main weakness is that a core part of the method specification, Eqs. (5) and (7), is internally inconsistent with respect to matrix dimensions, which makes the current text ambiguous enough to prevent reproduction and to obscure which of the two distillation losses is responsible for the reported ablation gains.","major_comments":[{"comment":"The matrix definitions are dimensionally inconsistent. Section III-C1 defines f_j^cam, f_j^teacher in R^{N×C}. In Eq. (5), f_j f_j^T is N×N, not C×C, so it cannot represent an inter-channel relationship matrix; the correct channel-correlation matrix is f_j^T f_j. In Eq. (7), f_j^T f_j is C×C, not N×N, so it cannot represent an inter-keypoint relationship matrix; the correct keypoint-correlation matrix is f_j f_j^T. Figure 6 and the text label the losses as inter-channel and inter-keypoint, but the formulas are swapped relative to those labels and to the stated dimensions. This makes the method specification ambiguous: if the implementation follows the equations literally, the loss named L_IC^bev is actually comparing keypoint correlations and L_IK^bev is comparing channel correlations, so the Table V ablation cannot be mapped to the described components. Please correct the equations or the stated dimensions/names and confirm which variant is used in the released code.","section":"Section III-C2 and III-C3, Eqs. (5) and (7)"},{"comment":"The overall loss in Eq. (10) is written as an unweighted sum, but no loss weights are reported for L_depth^R, L_bev^IC, or L_bev^IK. These auxiliary losses operate on very different scales (continuous relative depth values versus similarity matrices of size C×C or N×N), so equal weights are unlikely to be the actual setting. In addition, Section III-C1 samples N keypoints per target and slightly enlarges ground-truth boxes in BEV space, but neither the value of N nor the enlargement factor is reported in Section IV-B. These are free parameters of the method and are necessary for reproducing the ablation results in Tables III, V, and VIII. Please report the exact loss weights, N, the enlargement factor, and any normalization used in Eqs. (6) and (8).","section":"Section III-D and Section IV-B"},{"comment":"The adaptive depth reference is selected by minimizing the error between predicted and ground-truth depth at training time. This is legitimate for a training-only auxiliary loss, but the paper should state explicitly that this selection is not used at inference, and it should discuss the oracle nature of using ground-truth depth in the reference selection. Table IV compares the 'Smallest Error' policy with a 'Highest Conf' policy and shows a relatively small gap, which partially addresses the concern, but the paper would be strengthened by an additional sensitivity analysis, for example using a fixed reference point or a predicted-only confidence reference, and by reporting how the reference selection behaves over training iterations.","section":"Section III-B3, Eq. (2)"},{"comment":"The comparison with 'Naive Distill' uses dense BEV MSE, and the BEVDistill comparison uses foreground-guided dense distillation; neither controls for the keypoint sampling step. To substantiate the claim that relationship-based distillation is superior to direct feature alignment, please add an ablation that applies a point-wise L2 loss to the same N sampled keypoints without computing cross-channel or cross-keypoint correlations. Without this control, the gains attributed to inter-channel and inter-keypoint relationship modeling could also come from the keypoint sampling procedure or from the foreground masking itself.","section":"Section IV-D6, Table VIII"},{"comment":"For occluded, distant, or LiDAR-sparse objects, the foreground pixel set obtained by projecting LiDAR points inside the box can be empty or very small. The paper acknowledges the occlusion limitation in Section IV-F2, but it does not specify how Eq. (4), Eq. (6), and Eq. (8) are computed for degenerate cases, for example whether targets with fewer than a threshold number of pixels are skipped. The distance-range ablation in Table XII shows only modest gains at 30-60 m, so the handling of sparse foreground is empirically relevant. Please report the fraction of targets with valid foreground in the training set and state the exact handling rule for empty or near-empty target regions.","section":"Section III-B1 and Section IV-F2"}],"minor_comments":[{"comment":"The text states that TiGDistill-BEV 'consistently produced lower error metrics', but Table I shows that BEVDet mAOE increases from 52.3 to 55.4 and BEVDet4D mA VE increases from 33.1 to 34.1; please qualify this claim.","section":"Section IV-C1"},{"comment":"The '13.2% reduction in mA VE' compares BEVDet's 90.9 to 77.7, which is an absolute difference of 13.2 units rather than a percentage reduction, and it is not 'compared to BEVDepth' as written; please correct the description.","section":"Section IV-C1"},{"comment":"The text says PillarNext* gains 2.1% NDS, but Table VII shows 50.4 - 48.1 = 2.3; please correct this value.","section":"Section IV-D5, Table VII"},{"comment":"The number of depth bins is denoted K in Section III-A but D in Section III-B2; please use consistent notation and define the symbol \\hat{S}[k].","section":"Section III-B2, Eq. (1)"},{"comment":"The L2 norm is applied to sets of depth values; please write the loss as an explicit sum over pixels with clear indices so that the objective is unambiguous.","section":"Section III-B4, Eq. (4)"},{"comment":"Eq. (10) uses L_bev, but Eq. (9) defines the symbol L_bev^D; please align the notation.","section":"Section III-D, Eq. (10)"},{"comment":"The two distillation terms are not mathematically orthogonal; consider using 'complementary' instead of 'orthogonal'.","section":"Section III-C, Eq. (9)"},{"comment":"There are typos in the figures and captions: 'Target Iner-Geometry' in Figure 4, 'Chosely' in the Figure 4 diagram, and 'Innter-feature' in Figure 6; please proofread.","section":"Figures 4 and 6"},{"comment":"The sentence 'based on the the predicted continuous depth values' contains a duplicated 'the'.","section":"Section III-B3"},{"comment":"The sentence 'The codes is available at: ...' should read 'The code is available at: ...'.","section":"Abstract"},{"comment":"The BEVDet baseline values differ across rows (20.3/33.1, 30.5/37.8, and 29.8/37.9) although the caption says the configurations uniformly use ResNet-50 at 256×704; please clarify which baseline corresponds to each concurrent method.","section":"Table X"},{"comment":"Table VI reports SILog, but SILog is not defined in the evaluation metrics section; please add its definition.","section":"Section IV-A and Table VI"},{"comment":"MonoDETR is listed twice with the same title; please consolidate the duplicate reference.","section":"References [42] and [43]"},{"comment":"The sentence 'we utilize the freeze the teacher model' has a grammar error; please correct it.","section":"Section IV-B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of a computer vision journal and the empirical results are potentially valuable, but the dimensional inconsistency in Eqs. (5) and (7) and the missing training hyperparameters are serious reproducibility blockers. I did not verify the released code; the authors should be asked to confirm that the corrected equations match the implementation and to report the exact loss weights and keypoint count. The issues appear fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper reports a plausible, well-ablated distillation recipe for camera BEV detectors, and the gains look real, but the two central inner-feature equations are dimensionally inconsistent as written. A referee cannot reproduce the module from the manuscript until that is fixed.\n\nWhat is genuinely new: applying inner-depth relative supervision with an adaptive reference point, plus relationship-based BEV distillation on sampled foreground keypoints. The authors are honest about their inspirations ([45], [46]); the combination is not in prior work. The ablation work is thorough: per-component gains, teacher model comparison, small-object classes, distance ranges, and a matched comparison with BEVDistill. The reported val/test gains on nuScenes are consistently positive across baselines, and nothing in the numbers trips an alarm.\n\nThe soft spot is real, not cosmetic. In Section III-C, keypoint features are declared N×C. Equation (5) defines A = f f^T and says A is C×C; for an N×C matrix, f f^T is N×N, and the channel-correlation matrix is f^T f. Equation (7) defines B = f^T f and says B is N×N; that is backwards. So either the implementation follows the equations and the inter-channel and inter-keypoint losses are swapped relative to their names and ablations, or it follows the dimensional claims and the equations are wrong. Either way, the paper does not specify the method. This is a load-bearing flaw in a core contribution, not a footnote typo, and Table V cannot be mapped to the described components until it is resolved.\n\nOther issues are more standard: no error bars, loss weights and the number of keypoints N are not reported, and the SOTA comparison mixes implementations and backbones. The adaptive reference point uses ground-truth depth at training time; that is legitimate for a training loss, but sensitivity to that choice is only partially explored. The occlusion limitation is acknowledged in Section IV-F2.\n\nBottom line: the empirical story deserves serious referee time, and the paper is close to being a solid subfield contribution. But as submitted, the inconsistency in the core equations makes the method unreproducible. I would send it to review with a strong request to fix the transposition and release exact configs; I would not cite it as is.","headline":"Real gains on nuScenes, but the core inner-feature equations are dimensionally inconsistent as written; the method needs major revision before it is reproducible.","tokens_in":20979,"tokens_out":3775,"would_cite":false,"duration_ms":34521,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Distilling an object's internal depth relations and keypoint-feature similarities from a LiDAR or fusion teacher lifts a camera-only bird's-eye-view detector to 62.8% NDS on nuScenes.","keywords":["3D object detection","bird's-eye view","knowledge distillation","LiDAR-to-camera distillation","inner-geometry learning","depth supervision","BEV feature distillation","nuScenes"],"falsifier":"A direct test: on the nuScenes validation set, bucket objects by occlusion level and by the number of LiDAR points inside the ground-truth box; if TiGDistill-BEV's NDS improvement over BEVDepth is concentrated in well-observed buckets and vanishes for occluded or sparse objects, the foreground-projection premise is load-bearing. Alternatively, randomly drop a fraction of the foreground pixels used in the inner-depth loss during training; if the gain survives heavy dropout, the premise is not essential.","tokens_in":19891,"feed_emoji":"🚗","tokens_out":9517,"duration_ms":76797,"temperature":0.7,"pith_summary":"This paper tries to establish that a camera-only bird's-eye-view (BEV) 3D detector can be made substantially more accurate by distilling, from a LiDAR or fusion teacher, the internal geometry of each foreground object rather than only dense depth or dense BEV features. The proposed scheme, TiGDistill-BEV, adds two complementary supervisory signals: relative depth within each object computed against an adaptively chosen reference pixel, and the similarity structure of BEV keypoint features sampled inside each object, matched across channels and across keypoints. On nuScenes, applying this to BEVDepth raises validation NDS from 52.1 to 54.4 and test NDS from 58.9 to 62.8, with mAP from 49.1 to 53.9, surpassing prior distillation-based detectors. If the result holds, the active ingredient in LiDAR-to-camera distillation is not matching raw features but transferring the object's internal spatial and semantic relations.","feed_headline":"Camera-only detection hits 62.8 NDS with inner-geometry distillation","feed_subtitle":"LiDAR teachers transfer an object's relative depth and keypoint-feature relations to camera-only detectors.","key_machinery":"Two modules carry the method. Inner-depth supervision projects LiDAR points inside each ground-truth 3D box onto image planes to obtain foreground pixels, converts the student's categorical depth distribution into a continuous depth per pixel, picks the pixel with the smallest depth error as an adaptive reference, and applies an $L_2$ loss to the predicted-minus-reference depth differences. Inner-feature BEV distillation samples N uniform keypoints inside each enlarged BEV box, extracts their features from teacher and student BEV encoders, and matches two similarity matrices under $L_2$ loss: the $C\\times C$ inter-channel matrix and the $N\\times N$ inter-keypoint matrix of the sampled features. These similarity matrices are the central objects; they carry object-level geometric semantics without forcing cross-modal feature identity.","core_discovery":"The central claim is that target inner-geometry learning—supervising the camera student on relative depth inside each object and on inter-channel and inter-keypoint similarities of its BEV keypoint features—is what makes cross-modal distillation effective for multi-view BEV detection. The paper argues that absolute depth supervision ignores fine-grained within-object structure and that dense BEV feature imitation is hampered by the modality gap. By converting categorical depth predictions to continuous values and subtracting an adaptively chosen reference depth per object, and by distilling only the correlation matrices of sampled foreground keypoints instead of the features themselves, the student inherits LiDAR-like geometric understanding while keeping its own representational identity. The reported outcome is consistent gains across student baselines, backbones, temporal settings, small objects, and distance ranges, with the largest headline result of 62.8% NDS and 53.9% mAP on the nuScenes test set.","pith_inferences":["Beyond the paper: if relative object structure is the active signal, the same inner-geometry losses could transfer to other camera-only BEV tasks such as BEV map segmentation or occupancy prediction, where LiDAR teachers are available.","Beyond the paper: the adaptive reference point is selected using ground-truth depth error at training time; a learned or confidence-weighted reference might generalize better when student depth is systematically biased, a comparison the paper does not run.","Beyond the paper: because the stated motivation is the modality gap, the advantage over dense BEV distillation should grow as teacher and student architectures diverge; testing with deliberately mismatched backbones would isolate this effect.","Beyond the paper: the paper lists temporal inner-depth supervision as future work; aggregating the same object's interior points across frames could recover occluded foreground, provided multi-frame calibration is accurate."],"forward_implications":["BEVDepth, BEVDet, and BEVDet4D all improve on the nuScenes val set under a CenterPoint teacher, with NDS gains between 2.3 and 2.9 points.","On the nuScenes test set, BEVDepth with a BEVFusion teacher reaches 62.8% NDS and 53.9% mAP, outperforming BEVDistill by 3.4 NDS and 4.1 mAP under matched implementation.","Inner-feature distillation alone contributes more (+3.0 mAP, +2.3 NDS) than inner-depth supervision alone (+1.0 mAP, +0.9 NDS), and combining both gives +3.7 mAP and +3.0 NDS on the val ablation.","The improvement persists across image backbones (ResNet-18/50/101, Swin, ConvNeXt), in single-frame and multi-frame settings, for small object classes, and in the 30-60 meter range.","Teacher quality scales the transfer: using BEVFusion as teacher yields larger gains than LiDAR-only teachers CenterPoint and PillarNeXt."],"supporting_citations":[{"why":"Provides the BEVDepth camera-based student baseline with dense absolute depth supervision that TiGDistill-BEV extends.","marker":"[15]"},{"why":"BEVDistill is the dense foreground-guided BEV distillation baseline that TiGDistill-BEV compares against and outperforms.","marker":"[16]"},{"why":"CenterPoint supplies the LiDAR-only teacher used in the validation experiments.","marker":"[1]"},{"why":"BEVFusion supplies the LiDAR-camera fusion teacher used for the test-set state-of-the-art result.","marker":"[68]"},{"why":"LSS defines the camera-to-BEV projection mechanism that the student's predicted depth feeds into.","marker":"[22]"},{"why":"BEVDet is one of the student baselines used to show the method generalizes beyond BEVDepth.","marker":"[20]"},{"why":"BEVDet4D is the temporal student baseline used to show the method works in multi-frame settings.","marker":"[21]"},{"why":"nuScenes supplies the dataset and the NDS/mAP evaluation metrics for all reported results.","marker":"[69]"}],"fun_headline_variants":["Inner-geometry distillation lifts camera-only BEV to 62.8 NDS","TiGDistill-BEV: Teaching cameras LiDAR's inner geometry for 62.8 NDS","From LiDAR to cameras: Inner-geometry distillation hits 62.8 NDS","Distill inner-geometry, not just depth: 62.8 NDS for camera-only 3D","Bridging the LiDAR-camera gap with target inner-geometry distillation: 62.8 NDS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that projecting the LiDAR points inside a ground-truth 3D box onto image and BEV planes yields a reliable foreground mask for every training target, so the inner-depth and inner-feature losses supervise true object geometry rather than noisy, sparse, or occluded pixels.","fun_headline_variants_meta":{"raw":{"variants":["Inner-geometry distillation lifts camera-only BEV to 62.8 NDS","TiGDistill-BEV: Teaching cameras LiDAR's inner geometry for 62.8 NDS","From LiDAR to cameras: Inner-geometry distillation hits 62.8 NDS","Distill inner-geometry, not just depth: 62.8 NDS for camera-only 3D","Bridging the LiDAR-camera gap with target inner-geometry distillation: 62.8 NDS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001064,"raw_usage":{"total_tokens":4496,"prompt_tokens":1016,"completion_tokens":3480,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":3356}},"tokens_in":632,"tokens_out":3480,"duration_ms":23663,"temperature":1.0,"reasoning_tokens":3356,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:05:52.625066+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test: on the nuScenes validation set, bucket objects by occlusion level and by the number of LiDAR points inside the ground-truth box; if TiGDistill-BEV's NDS improvement over BEVDepth is concentrated in well-observed buckets and vanishes for occluded or sparse objects, the foreground-projection premise is load-bearing. Alternatively, randomly drop a fraction of the foreground pixels used in the inner-depth loss during training; if the gain survives heavy dropout, the premise is not essential.","supporting_citations":[{"cited_title":"Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation,","cited_arxiv_id":null,"evidence_quote":"BEVFusion supplies the LiDAR-camera fusion teacher used for the test-set state-of-the-art result."}],"review_version":1}