Pith. sign in

REVIEW 5 major objections 5 minor 31 references

A lightweight radar-camera depth estimator can match a heavy teacher's accuracy by distilling the teacher's saliency maps and depth distributions, cutting parameters by 29.7% and MAE by 7.97%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

XD-RCDepth uses explainability-aligned and depth-distribution distillation to shrink a radar-camera depth model by 29.7% parameters while improving MAE by about 8%.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Solid engineering paper with a novel explainability-distillation idea; the reported gains are plausible but under-reported details keep the exact numbers from being fully trustworthy. the 5 major comments →

arxiv 2510.13565 v3 pith:F57BPKOC submitted 2025-10-15 cs.CV

XD-RCDepth: Lightweight Radar-Camera Depth Estimation with Explainability-Aligned and Distribution-Aware Distillation

Huawei Sun , Zixu Wang , Xiangyuan Peng , Julius Ott , Georg Stettinger , Lorenzo Servadei , Robert Wille This is my paper
classification cs.CV
keywords depth estimationradar-camera fusionknowledge distillationexplainable AIGrad-CAMFiLMlightweight networkautonomous driving
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that depth estimation from radar-camera fusion can be made lightweight without sacrificing accuracy, by transferring knowledge from a large teacher to a small student in two underused ways: aligning the teacher's saliency maps (via Grad-CAM) and matching the teacher's soft depth-bin distributions. If true, this would make accurate depth perception feasible on embedded platforms in autonomous vehicles and give engineers a way to keep models interpretable. The authors report a 29.7% parameter reduction relative to the lightweight LiRCDepth baseline, with a 7.97% lower MAE than training the same student without distillation on nuScenes at the 80m range, and consistent gains on ZJU-4DRadarCam. They also show that either distillation loss alone helps, and that together they produce sharper saliency maps and cleaner depth boundaries.

Core claim

The central discovery is that two complementary distillation objectives — an explainability-aligned loss that aligns Grad-CAM saliency maps of selected intermediate features between teacher and student, and a depth-distribution loss that recasts continuous depth regression as a soft classification over discretized depth bins — allow a MobileNetV2-based student with FiLM fusion and a point-wise DASPP decoder to recover accuracy lost during compression. Applied to the LiRCDepth teacher-student setup, these losses produce consistent MAE reductions, and in the XD-RCDepth model they deliver a 7.97% MAE improvement over the non-distilled student at 80 meters while using 29.7% fewer parameters than

What carries the argument

The key machinery is the pair of distillation losses: X-KD, a Grad-CAM-based explainability-aligned saliency map distillation that computes per-layer cosine distance between ℓ2-normalized teacher and student saliency maps (detaching the teacher), and D2-KD, a depth-distribution distillation that converts each pixel's teacher and student depth into softmax distributions over B depth bins (with temperature τ) and minimizes the forward KL divergence from teacher to student. These are supported by a FiLM fusion module that modulates image features with per-channel scale and shift predicted from radar features, and a point-wise DASPP decoder that widens receptive field with negligible parameters.

Load-bearing premise

The 7.97% MAE improvement over direct training assumes the non-distilled student is trained with the same full schedule and hyperparameters as the distilled one; the paper does not specify the training epochs, optimizer, or learning-rate schedule for either model.

What would settle it

Retrain the w/o KD student on nuScenes using exactly the same epoch count, optimizer settings, and learning-rate schedule as the XD2-KD run; if the MAE at 80m does not remain roughly 7.97% higher than the distilled model, the claimed gain conflates distillation with optimization effects.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the claims hold, the 29.7% parameter cut with comparable accuracy means radar-camera depth estimation can run at ~15 fps on embedded GPUs, making onboard deployment practical.
  • Explainability-aligned distillation gives engineers a student that is more interpretable (saliency maps closer to the teacher's), which could matter for safety-critical validation.
  • The depth-distribution distillation suggests that modeling depth as a soft categorical distribution over bins — rather than direct regression — is a more transferable target, a principle that may extend to other dense regression tasks.
  • The gains are shown on both 3D (nuScenes) and 4D (ZJU-4DRadarCam) radar, suggesting the method is not tied to a particular radar representation.
  • Rebasing the student on the same teacher (CaFNet) with two different backbones shows the distillation recipes are portable across student architectures.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reported 7.97% gain assumes the 'w/o KD' baseline is trained to convergence with the same schedule; if the baseline were undertrained, a direct replication could show a smaller gap.
  • The X-KD loss's use of Grad-CAM with the image-level mean depth as the scalar objective means the saliency alignment transfers only coarse, scene-level attention; an extension would be to use dense pixel-wise attribution targets to align fine-grained structure.
  • The binning plus KL-divergence formulation is reminiscent of ordinal regression; one could test adaptive binning or unequal bin widths to push accuracy further, something the paper does not explore.
  • The mutual benefit of the two losses hints at a regularizing effect: saliency alignment may prevent the student from attending to spurious features while distribution matching calibrates its depth outputs.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper presents XD-RCDepth, a lightweight radar-camera depth estimation network. The student model uses MobileNetV2 backbones for image and radar streams, a compact FiLM fusion module, and a point-wise DASPP decoder. Training combines a standard depth MAE loss with two knowledge-distillation losses: X-KD aligns Grad-CAM saliency maps of selected teacher/student layers, and D2-KD matches soft per-pixel distributions over discretized depth bins. Against the LiRCDepth baseline, the architecture reduces parameters by 29.7%, and the full method is reported to improve MAE by 7.97% over its non-distilled counterpart on nuScenes, with additional results on ZJU-4DRadarCam. The authors claim competitive accuracy with heavy-weight models and real-time efficiency.

Significance. If the results hold, the paper makes a useful practical contribution: it shows that a smaller radar-camera depth network can retain accuracy through distillation, and it extends explainability-aware distillation from classification to dense prediction. The code release, the ablation of individual components, and the use of an external teacher baseline are strengths. However, the quantitative claims are currently supported only by single-run numbers with several missing experimental details, so the precise magnitude of the reported gains is not yet firmly established.

major comments (5)
  1. [Section 3.1/3.2] The control condition for the central 7.97% MAE claim is not fully specified. Section 3.1 reports only GPU, batch size, and loss weights; no optimizer, learning rate, schedule, or number of epochs is given. Table 1 compares XD-RCDepth (w/o KD) with XD-RCDepth (XD2-KD), but if the non-distilled model is trained with a shorter or different schedule, the gain could come from optimization rather than distillation. State the exact training protocol and confirm that all models used in Tables 1, 3, and 4 are trained with the same budget and hyperparameters.
  2. [Tables 1 and 4] All quantitative results are single runs without error bars. The ablation differences are small: in Table 4, MAE drops from 2.232 to 2.054, while X-KD alone (2.114) and D2-KD alone (2.132) differ by only 0.018, and the incremental gain of adding D2-KD to X-KD is 0.060. Such differences could be within run-to-run noise. Report mean and standard deviation over at least three seeds, or provide a statistical test, for the central comparisons.
  3. [Section 2.3 / Section 3.1] D2-KD depends on hyperparameters B, [d_min, d_max], and tau, but the experiments never state their values. Equations (7)-(8) and the qualitative description are not enough to reproduce the method. Similarly, X-KD is said to distill from 'three layers (image encoder, radar encoder, decoder) at 1/16 scale' without identifying the actual layer names. Please report these settings.
  4. [Section 3.2 and Table 1] The real-time claim is internally inconsistent. The text states the model 'achieves about 15 frames per second,' but Table 1 reports runtime 0.015 s for XD-RCDepth, which is about 67 FPS. It is unclear which is correct and under what measurement conditions (GPU, batch size, warm-up, whether preprocessing is included). Please reconcile and specify the benchmark protocol.
  5. [Table 2] On ZJU-4DRadarCam, the paper reports only RadarNet, CaFNet, and the proposed method. No lightweight baseline such as LiRCDepth is included, and only four metrics are shown. This makes it difficult to support the claim that the method is competitive with the leading lightweight baseline on both datasets. Add the corresponding baseline and full metric set, or temper the claim.
minor comments (5)
  1. [Section 2.1] The Point-wise DASPP block is described only verbally. Please provide a precise configuration table or pseudocode: number of branches, dilation rates, kernel sizes, and how the branches are aggregated.
  2. [Equation (4)] The epsilon in the denominator is undefined. State its value.
  3. [Table 1 caption] The notation 'XD2-KD' in Table 1 is used for the proposed distillation applied to LiRCDepth, but the caption only says 'distillation approaches proposed in this paper.' Clarify the relationship between 'KD' and 'XD2-KD' rows.
  4. [Section 3.4] The sentence 'Increasing 0.4M parameters, resulting in an improvement of the MAE by 2.3%' is ambiguous. State explicitly that adding the point-wise DASPP increases parameters from 8.45M to 8.89M and improves MAE from 2.285 to 2.232.
  5. [Figures 2 and 3] The qualitative figures would be easier to evaluate with colorbars and explicit labels for the rows/columns, especially the saliency maps in Fig. 3, where it is not immediately clear which image or layer is shown.

Circularity Check

0 steps flagged

No significant circularity; distillation gains are supported by a direct w/o-KD control and independent test-set evaluation.

full rationale

XD-RCDepth's central claims—parameter reduction and the ~8% MAE improvement from distillation—are established by direct empirical comparison. The student with and without X-KD/D2-KD is evaluated under the same architecture: Table 1 reports the w/o-KD control (MAE 2.232 at 80 m) versus XD2-KD (MAE 2.054), so the gain is not a fitted parameter renamed as a prediction. The teacher CaFNet is self-cited, but it is used as a fixed distillation target and as an external performance reference; the distillation losses are additive regularizers on top of the depth loss, and success is measured on a separate test set. No equation in Section 2 defines the reported MAE in terms of the proposed losses by construction; X-KD and D2-KD are objective terms, not derivation outputs. The 29.7% parameter reduction is an architectural comparison (8.89M vs. 12.65M parameters). The main substantive concern is whether the w/o-KD baseline received the same training schedule and hyperparameters, which affects the magnitude of the gain but is a question of experimental control, not circularity. The self-citations to the baseline and teacher are real references for comparison rather than load-bearing justifications of the method's correctness.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claims rest on several chosen hyperparameters (loss weights, temperature, bin count, layer choices) and on domain assumptions about the sufficiency of FiLM fusion and of Grad-CAM-based saliency transfer. None of these are fitted to data in a circular way, but they are unstated degrees of freedom.

free parameters (6)
  • loss weights lambda_1, lambda_2, lambda_3 = 1.0, 0.5, 0.5
    Set by hand (Sec 3.1) without reported tuning; the ablation in Table 4 only tests presence/absence, not weight values.
  • temperature tau in D2-KD
    Value not reported; affects the softness of the target distribution in Eq. 7.
  • number of depth bins B
    Not stated in text; controls granularity of depth distribution in Eq. 6.
  • depth range [d_min, d_max] for bins
    Needed to define bin centers; not specified in the paper.
  • distillation layers L = image encoder, radar encoder, decoder at 1/16 scale
    Chosen (Sec 3.1) without ablation; the method's sensitivity to layer selection is unknown.
  • Grad-CAM target = mean predicted depth
    Chosen as the scalar objective in Eq. 3; the paper acknowledges future work on alternative targets (Sec 4).
axioms (5)
  • domain assumption Grad-CAM saliency maps are meaningful and transferable supervision for dense prediction.
    The X-KD loss (Eq. 5) assumes that matching teacher and student intermediate saliency improves the student; no theoretical justification given.
  • domain assumption Soft classification over depth bins is an effective surrogate for regression.
    Motivated by DORN [22]; the paper assumes the KL loss is a good training signal.
  • domain assumption Radar features can be fused with image features via per-channel affine modulation (FiLM).
    Used in Eq. 1-2; validated only via ablation Table 3.
  • domain assumption The teacher (CaFNet) provides accurate saliency and depth targets.
    Self-cited [12]; no independent verification of teacher quality beyond its metrics in Table 1.
  • domain assumption Empirical gains on nuScenes and ZJU-4DRadarCam generalize to other radar-camera settings.
    Only two datasets tested; no claims of universality, but this is assumed for practical use.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of XD-RCDepth: Lightweight Radar-Camera Depth Estimation with Explainability-Aligned and Distribution-Aware Distillation." pith.science (2026). https://pith.science/paper/F57BPKOC

@misc{pith2026251013565,
  author       = {Pith},
  title        = {Pith review of: XD-RCDepth: Lightweight Radar-Camera Depth Estimation with Explainability-Aligned and Distribution-Aware Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F57BPKOC}},
  note         = {Machine review of arXiv:2510.13565}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Depth estimation remains central to autonomous driving, and radar-camera fusion offers robustness in adverse conditions by providing complementary geometric cues. In this paper, we present XD-RCDepth, a lightweight architecture that reduces the parameters by 29.7% relative to the state-of-the-art lightweight baseline while maintaining comparable accuracy. To preserve performance under compression and enhance interpretability, we introduce two knowledge-distillation strategies: an explainability-aligned distillation that transfers the teacher's saliency structure to the student, and a depth-distribution distillation that recasts depth regression as soft classification over discretized bins. Together, these components reduce the MAE compared with direct training with 7.97% and deliver competitive accuracy with real-time efficiency on nuScenes and ZJU-4DRadarCam datasets. Code: https://github.com/harborsarah/XD_RCDepth

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 3 linked inside Pith

  1. [1]

    The task typically involves predicting a dense depth map either from an RGB image [1–4] or from an RGB image augmented with a sparse LiDAR depth map [5–7]

    INTRODUCTION Despite rapid progress in autonomous driving, depth estima- tion remains critical and comparatively underexplored. The task typically involves predicting a dense depth map either from an RGB image [1–4] or from an RGB image augmented with a sparse LiDAR depth map [5–7]. Purely camera-based approaches are ill-posed since RGB provides no direct...

    Pith/arXiv arXiv 2025

  2. [2]

    After- wards, the explainability-aligned CAM distillation and the depth distribution distillation methods are presented

    METHODOLOGY This section first introduces the model architecture. After- wards, the explainability-aligned CAM distillation and the depth distribution distillation methods are presented. At last, the loss functions are summarized. 2.1. Model Architecture As illustrated in Fig.1, to remain comparable with LiR- CDepth [16], we adopt CaFNet [12] as the teach...

  3. [3]

    Datasets and Implementation Details We experiment on nuScenes [14] and ZJU-4DRadarCam [11] datasets to validate our approach under both 3D- and 4D-radar regimes

    EXPERIMENTS 3.1. Datasets and Implementation Details We experiment on nuScenes [14] and ZJU-4DRadarCam [11] datasets to validate our approach under both 3D- and 4D-radar regimes. For nuScenes, we construct the supervision depth Dsusing the same accumulation procedure as in [12]. For ZJU–4DRadarCam, we directly use the provided dense depth map asDs. All mo...

  4. [4]

    Study of fusion methods and point-wise DASPP

    CONCLUSION We present XD-RCDepth, a lightweight radar–camera depth estimation framework that combines an efficient FiLM fusion Table 3. Study of fusion methods and point-wise DASPP. Fusion Method Params↓MAE↓RMSE↓AbsRel↓δ 1 ↑ add 8.74M 2.248 4.903 0.115 0.886 concatenate 10.94M 2.208 4.802 0.114 0.888 attention 9.48M 2.266 4.901 0.115 0.885 FiLM 8.89M2.232...

  5. [5]

    ACKNOWLEDGEMENT Research leading to these results has received funding from the EU ECSEL Joint Undertaking under grant agreement n° 101007326 (project AI4CSM) and from the partner national funding authorities the German Ministry of Education and Research (BMBF)

  6. [6]

    From big to small: Multi-scale local planar guidance for monocular depth estimation,

    Jin Han Lee et al., “From big to small: Multi-scale local planar guidance for monocular depth estimation,”arXiv preprint arXiv:1907.10326, 2019

    Pith/arXiv arXiv 1907

  7. [7]

    Iebins: Iterative elastic bins for monocular depth estimation,

    Shuwei Shao et al., “Iebins: Iterative elastic bins for monocular depth estimation,” inAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  8. [8]

    Unidepth: Universal monocular metric depth estimation,

    Luigi others Piccinelli, “Unidepth: Universal monocular metric depth estimation,” inIEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), 2024

    2024

  9. [9]

    P3depth: Monocular depth esti- mation with a piecewise planarity prior,

    Vaishakh Patil et al., “P3depth: Monocular depth esti- mation with a piecewise planarity prior,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1610–1621

    2022

  10. [10]

    Bilateral propagation network for depth completion,

    Jie Tang et al., “Bilateral propagation network for depth completion,” inProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 9763–9772

    2024

  11. [11]

    Improving depth completion via depth feature upsampling,

    Yufei Wang et al., “Improving depth completion via depth feature upsampling,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 21104–21113

    2024

  12. [12]

    Lrru: Long-short range recurrent up- dating networks for depth completion,

    Yufei Wang et al., “Lrru: Long-short range recurrent up- dating networks for depth completion,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 9422–9432

    2023

  13. [13]

    4d mmwave radar for sensing enhancement in adverse environments: Advances and challenges,

    Xiangyuan Peng et al., “4d mmwave radar for sensing enhancement in adverse environments: Advances and challenges,” 2025

    2025

  14. [14]

    Depth estimation from cam- era image and mmwave radar point cloud,

    Akash Deep Singh et al., “Depth estimation from cam- era image and mmwave radar point cloud,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 9275–9285

    2023

  15. [15]

    Get-up: Geometric-aware depth estimation with radar points upsampling,

    Huawei Sun et al., “Get-up: Geometric-aware depth estimation with radar points upsampling,” in2025 IEEE/CVF Winter Conference on Applications of Com- puter Vision (WACV), 2025, pp. 1850–1860

    2025

  16. [16]

    Radarcam-depth: Radar-camera fusion for depth estimation with learned metric scale,

    Han Li et al., “Radarcam-depth: Radar-camera fusion for depth estimation with learned metric scale,” 2024

    2024

  17. [17]

    Cafnet: A confidence-driven frame- work for radar camera depth estimation,

    Huawei Sun et al., “Cafnet: A confidence-driven frame- work for radar camera depth estimation,” 2024

    2024

  18. [18]

    Tride: A text-assisted radar-image weather-aware fusion network for depth estimation,

    Huawei Sun et al., “Tride: A text-assisted radar-image weather-aware fusion network for depth estimation,” 2025

    2025

  19. [19]

    nuscenes: A multimodal dataset for autonomous driving,

    Holger Caesar et al., “nuscenes: A multimodal dataset for autonomous driving,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11621–11631

    2020

  20. [20]

    Radar-camera pixel depth asso- ciation for depth completion,

    Yunfei Long et al., “Radar-camera pixel depth asso- ciation for depth completion,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 12507–12516

    2021

  21. [21]

    Lircdepth: Lightweight radar- camera depth estimation via knowledge distillation and uncertainty guidance,

    Huawei Sun et al., “Lircdepth: Lightweight radar- camera depth estimation via knowledge distillation and uncertainty guidance,” inICASSP 2025 - 2025 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

    2025

  22. [22]

    Distill- ing the knowledge in a neural network,

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean, “Distill- ing the knowledge in a neural network,”arXiv preprint arXiv:1503.02531, 2015

    Pith/arXiv arXiv 2015

  23. [23]

    Explainable artificial intelligence (xai): What we know and what is left to attain trustworthy ar- tificial intelligence,

    Sajid Ali et al., “Explainable artificial intelligence (xai): What we know and what is left to attain trustworthy ar- tificial intelligence,”Information Fusion, vol. 99, pp. 101805, 2023

    2023

  24. [24]

    Utilizing ex- plainable ai for improving the performance of neural networks,

    Huawei Sun, Lorenzo Servadei, et al., “Utilizing ex- plainable ai for improving the performance of neural networks,” 2022

    2022

  25. [25]

    Learning interpretation with explainable knowledge distillation,

    Raed Alharbi, Minh N. Vu, and My T. Thai, “Learning interpretation with explainable knowledge distillation,” 2021

    2021

  26. [26]

    Good teachers explain: Explanation-enhanced knowledge distillation,

    Amin Parchami-Araghi et al., “Good teachers explain: Explanation-enhanced knowledge distillation,” 2024

    2024

  27. [27]

    Deep ordinal regression network for monocular depth estimation,

    Huan Fu et al., “Deep ordinal regression network for monocular depth estimation,” inProceedings of the IEEE conference on computer vision and pattern recog- nition, 2018, pp. 2002–2011

    2018

  28. [28]

    Deep residual learning for image recognition,

    Kaiming He et al., “Deep residual learning for image recognition,” 2015

    2015

  29. [29]

    Mobilenetv2: Inverted residuals and linear bottlenecks,

    Mark Sandler et al., “Mobilenetv2: Inverted residuals and linear bottlenecks,” 2019

    2019

  30. [30]

    Deeplab: Semantic image segmentation with deep convolutional nets, atrous con- volution, and fully connected crfs,

    Liang-Chieh Chen et al., “Deeplab: Semantic image segmentation with deep convolutional nets, atrous con- volution, and fully connected crfs,” 2017

    2017

  31. [31]

    Grad-cam: Visual ex- planations from deep networks via gradient-based local- ization,

    Ramprasaath R. Selvaraju et al., “Grad-cam: Visual ex- planations from deep networks via gradient-based local- ization,”International Journal of Computer Vision, vol. 128, no. 2, pp. 336–359, Oct. 2019

    2019

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.