LeAD-M3D: Leveraging Asymmetric Distillation for Real-Time Monocular 3D Detection¶

Conference: ECCV 2026
arXiv: 2512.05663
Code: https://deepscenario.github.io/LeAD-M3D/
Area: Autonomous Driving / 3D Vision
Keywords: Monocular 3D Detection, Knowledge Distillation, Real-Time Inference, MixUp Denoising, YOLOv10

TL;DR¶

LeAD-M3D integrates three core components on top of an image-only YOLOv10-M3D baseline: Asymmetric MixUp Denoising Distillation (A2D2), a 3D-aware Consistent Matcher (CM3D) that incorporates 3D overlap into matching scores, and Confidence Gated Inference (CGI3D) that runs the 3D regression head only on high-confidence regions. Without relying on LiDAR, stereo images, or geometric priors, it achieves state-of-the-art accuracy on KITTI, Waymo, and Rope3D, while running up to 3.6 times faster than models of comparable accuracy and exceeding 60 FPS under TensorRT for its largest variant.

Background & Motivation¶

Monocular 3D detection (M3D) aims to recover the 3D position, orientation, and size of objects from a single RGB image. Under this setting, depth ambiguity remains an inevitable and core source of error, which is further amplified when the camera has non-zero roll/pitch angles (e.g., roadside cameras in a top-down view). To compensate for the missing depth information, mainstream approaches generally follow one of two directions: one supervises depth during training using extra modalities or data (such as LiDAR, stereo images, object shapes, temporal sequences, or ground planes), whereas the other injects geometric priors into the model (typically perspective assumptions like "2D height is inversely proportional to depth") to regularize depth. The limitation of the former is that LiDAR data is often inaccessible for drones and certain road scenarios, and multi-modal pipelines are heavy and complicated. The issue with the latter is that the perspective assumption holds only for vehicle-mounted front-view cameras where the pitch/roll angles are close to zero, rendering the model brittle to viewpoint changes and strictly restricting its deployment. Truly cross-view and cross-dataset generalizability is only possessed by methods that rely solely on 3D bounding box supervision without extra modalities or hard-coded priors, which is the exact category targeted by this paper.

Within this category, runtime efficiency is another dimension that has been sacrificed for a long time. Recent M3D works widely push detection accuracy to the extreme while ignoring inference speed, causing difficulties in deployment. Knowledge distillation (KD) is a mature paradigm to achieve both accuracy and efficiency by transferring knowledge from a large, slow teacher to a small, fast student. However, to create the necessary "asymmetry" where the teacher is stronger than the student, existing M3D distillation frameworks almost exclusively inject LiDAR data into the teacher—thereby reintroducing modal dependency. Furthermore, standard KD losses align all features indiscriminately, ignoring whether the teacher's prediction is accurate or how much a particular feature channel contributes to the final depth output.

The key insight of this paper is that asymmetry does not necessarily need to be created by "giving the teacher more information." Conversely, "giving the student a harder input" serves the same purpose. By showing the teacher two clean images while feeding the student a noisy, Mixed-Up version of these images, the student is forced to reconstruct the teacher's clean depth features under corrupted inputs, thereby reformulating the distillation as a denoising task using only images. The core idea is to reconstruct M3D distillation into "asymmetric MixUp denoising": the teacher processes clean images while the student processes MixUp images, with alignment achieved via a depth feature loss weighted by teacher prediction quality and channel importance. Together with a 3D-aware Consistent Matcher (CM3D) to ensure reliable teacher-student pairing, and Confidence Gated Inference (CGI3D) to restrict 3D regression only to high-confidence regions, the framework simultaneously achieves state-of-the-art accuracy and real-time inference without using LiDAR or geometric priors.

Method¶

Overall Architecture¶

The backbone of LeAD-M3D is YOLOv10-M3D, which is built by appending standard 3D detection heads (representing 3D offset, 3D dimension, orientation, depth, and depth uncertainty), a post-processing module, and a 3D loss to YOLOv10. This creates an inherently efficient, NMS-free real-time M3D baseline. Following YOLO convention, the model family is scaled into five variants (N/S/M/B/X), where the largest X-variant is first pre-trained as the teacher without using A2D2. Once frozen, it guides the distillation of student models of any scale.

During training, a pair of clean images is fed into the frozen teacher, while their MixUp counterpart is forwarded to the student. Both models share the same meta-architecture (though the student is smaller) and generate their respective predictions. CM3D matches each ground-truth object with the optimal predictions from both the teacher and student sides, thereby establishing teacher-student feature pairs for identical objects. A2D2 executes weighted alignment distillation exclusively on the instance depth features output by the depth head instead of general backbone activations. During inference, the teacher and MixUp operation are discarded. The student uses CGI3D to execute 2D/3D regression heads only on top-k high-confidence locations selected by the classification head. The three components govern different stages: A2D2 handles how to transfer the teacher's depth inference capability to the student, CM3D ensures reliable teacher-student and supervision matching, and CGI3D avoids wasting computation on background areas during inference.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Two Clean Images + Their MixUp Image"] --> T["Teacher (X-scale, Frozen)<br/>Processes Clean Images → Depth Features F^T"]
    A --> S["Student (N~X)<br/>Processes MixUp Image → Depth Features F^S"]
    T --> M["3D-Aware Consistent Matcher CM3D<br/>Pairs via Class + 2D IoU + 3D MGIoU"]
    S --> M
    M --> D["Asymmetric Denoising Distillation A2D2<br/>Quality & Importance-Weighted Depth Feature Loss"]
    D -->|Discard Teacher and MixUp during Inference| G["Confidence Gated Inference CGI3D<br/>Runs 3D regression only on top-k high-confidence areas"]
    G --> O["3D Box + Class"]

Key Designs¶

1. A2D2: Leveraging MixUp denoising to create asymmetry and distill depth inference into the student

Existing M3D distillation methods simplify the teacher's task and create a representation gap by feeding LiDAR to the teacher, at the cost of reintroducing modal dependency. A2D2 takes the opposite route: instead of adding information to the teacher, it "increases the difficulty" for the student. It mixes up two images at the pixel level to form a single input for the student, requiring the student to detect all objects from both images simultaneously. Compared to other augmentation methods, the key benefit of MixUp is that it retains the complete space range of all objects in both images while leaving geometric quantities like projection centers, depths, dimensions, and orientations unchanged. Consequently, no ground-truth annotations are lost during augmentation, naturally forming a denoising task to eliminate MixUp artifacts in the feature space. Because the teacher processes the corresponding clean images, the student is forced to reconstruct the teacher's clean depth features under corrupted inputs, simultaneously creating the asymmetry and the denoising objective without relying on LiDAR. This allows the teacher and student to share the same architecture, greatly reducing complexity. The authors verified that MixUp creates the largest depth gap: on the augmented KITTI validation set, the teacher's median depth error rises from 59 cm with clean images to 75 cm with MixUp images (whereas RandAugment/CutMix only yield 62/66 cm); a larger gap leads to higher distillation gains. Additionally, selecting the instance depth features of the depth head rather than generic backbone features for distillation was shown via ablation to be crucial: replacing them with backbone features dropped Moderate AP by 0.35%, indicating that depth features are a better target for 3D tasks.

2. Quality and Channel Importance Weighted Depth Feature Loss: Avoid treating all features equally

Standard KD losses align all features uniformly, ignoring the accuracy of the teacher's predictions and the contribution of individual channels to the final depth. A2D2 addresses these issues with two weight terms. The first is teacher quality weight: if the teacher predicts an object's depth inaccurately, aligning student features with it will only introduce noise. Therefore, the loss is discounted based on the relative depth error. Given the ground-truth depth \(z_i\) and the teacher's predicted depth \(\hat z_i^{\rm T}\), the quality metric is defined as:

\[\eta_i = \frac{z_i}{\max\!\bigl(\lvert z_i - \hat z_i^{\rm T}\rvert,\ \epsilon\bigr)},\qquad \epsilon=0.1\]

Using "relative" error (with \(z_i\) in the numerator) instead of absolute error prevents close-range objects (which naturally have smaller errors) from overwhelming the loss, making weights between close and distant objects comparable after normalization. The second is channel importance weight: since depth is regressed from depth features using the teacher's depth-head weights \(W^{\rm T}\), channels with larger absolute weights have a greater impact on the final depth. Thus, the normalized channel weight \(\omega_q=\lvert W_q^{\rm T}\rvert / \sum_{q'}\lvert W_{q'}^{\rm T}\rvert\) is used as the importance factor. The final loss is a weighted L1 loss on the teacher-student instance depth features paired by CM3D, scaled by both weights: \(\mathcal{L}^{\text{distill}}=\frac{1}{|B(\mathbf I)|}\sum_i\sum_q \omega_q\,\eta_i\,\lvert\mathcal F^{\rm T}_{i,q}-\mathcal F^{\rm S}_{i,q}\rvert\). Ablation studies show that removing the quality metric drops Moderate AP by 0.37%, while removing the importance metric drops it by 0.61%, proving both are effective. Offline KD (with a frozen teacher) is utilized; the authors found that switching to online self-distillation dropped AP by 0.85%, because the strongest distillation targets only emerge in the late training stages when the learning rate decays and depth estimation stabilizes.

3. CM3D: Injecting 3D overlap into matching scores for reliable supervision

Reliable pairing between predictions and ground-truth objects is vital for both standard supervised training and A2D2. Matching noise can lead to misalignments between teacher-student pairs, destabilizing training especially under MixUp where multiple 2D projections heavily overlap. The matching strategy of the YOLOv10-M3D baseline operates solely in 2D, ranking pairs with \(s_{i,j}^{\text{2D}}=\hat p_{i,c_j}^{\alpha}\,\operatorname{IoU}(\hat{\mathbf b}_i^{\text{2D}},\mathbf b_j^{\text{2D}})^{\beta}\) (a weighted product of class probability and 2D IoU) to select the top-k predictions. CM3D elevates this matching to 3D by multiplying a term representing the 3D overlap between the predicted and ground-truth bounding boxes:

\[s_{i,j}^{\text{2D/3D}} = s_{i,j}^{\text{2D}}\cdot \operatorname{MGIoU}(\hat{\mathbf b}_i^{\text{3D}},\mathbf b_j^{\text{3D}})^{\gamma}\]

Instead of standard 3D IoU, Marginalized Generalized IoU (MGIoU) is intentionally used. Standard IoU remains constantly 0 when two bounding boxes do not overlap, causing gradient vanishing, which is common during early training or for small objects. MGIoU projects two 3D shapes onto a set of unique normal vectors, computes the 1D generalized IoU for each direction, and marginalizes them. This provides a meaningful overlap proxy even for completely non-overlapping boxes, while combining position, size, and orientation in a scale-invariant manner (unlike corner-based losses). This design allows the two terms to complement each other: when 3D estimation is coarse, the 2D term stabilizes training, and as accuracy improves, the 3D term disambiguates overlapping predictions. Compared to static anchor matching commonly used in many M3D methods, this dynamic 2D/3D scoring provides significantly better disambiguation in crowded scenes and MixUp configurations. The computational cost is minimal, with MGIoU increasing training time by less than 2%.

4. CGI3D: Gating 3D regression to high-confidence regions to save computation

Monocular 3D detectors typically evaluate regression heads densely across the entire feature map, wasting computational power on background areas. CGI3D simplifies this: it forwards only the classification head densely over the entire feature map, selects the top-k locations based on class confidence, and then crops \(3\times3\) local patches around these locations. The 2D/3D regression heads are then evaluated sparsely only on these selected patches. The key to achieving "zero accuracy loss and identical outputs as dense evaluation" is that the effective receptive field of the regression heads is exactly \(3\times3\) (consisting of one \(3\times3\) convolution followed by two \(1\times1\) convolutions). Consequently, a \(3\times3\) patch contains all necessary local context for regression at each location, which is simpler than the \(7\times7\) grids used in RoI-Align and eliminates bilinear interpolation. For simplicity, dense computation is retained during training, and the sparse patch regression is only activated during inference. Experimentally, adding CGI3D to the N-scale model halves runtime and reduces FLOPs by ~75% with nearly identical AP. Further reducing the regression head channels from 128 to 64 saves another 0.7 ms and cuts FLOPs by ~40% while maintaining accuracy.

Loss & Training¶

The teacher model (LeAD-M3D X w/o A2D2) is first trained with standard supervised losses—classification loss \(\mathcal L^{\text{cls}}\), 2D box loss \(\mathcal L^{\text{2D}}\), and 3D box loss \(\mathcal L^{\text{3D}}\)—and then frozen. During the distillation phase, the student model of any scale is trained using a total loss that adds the distillation loss to the three supervised terms: \(\mathcal L=\mathcal L^{\text{cls}}+\mathcal L^{\text{2D}}+\mathcal L^{\text{3D}}+\mathcal L^{\text{distill}}\). Optimization is performed using Adam with an initial learning rate of 0.001, a weight decay of 0.0005, and a cosine learning rate scheduler with a 3-epoch warmup. The CM3D hyperparameters are set to \(\alpha=0.5,\beta=1.0,\gamma=1.0\). All experiments are executed on a single NVIDIA RTX 8000 GPU: teacher training takes approximately 34 hours, while student training peaks at approximately 60 hours for the X-scale model on the KITTI dataset.

Key Experimental Results¶

Main Results¶

Comparison with lightweight (<30M parameters) M3D methods on the KITTI test set, Car category, AP0.7 (3D|R40):

Method	Extra Data	Params (M)	Time (ms)	Easy	Mod.	Hard
MonoNeRD	LiDAR	6.6	1380.3	22.75	17.13	15.63
MonoLSS	—	21.5	20.2	26.11	19.15	16.94
LeAD-M3D S (Ours)	—	10.1	10.2	27.28	18.87	16.37
LeAD-M3D B (Ours)	—	24.9	13.9	29.10	20.17	18.34

Remarkably, LeAD-M3D S outperforms MonoNeRD in accuracy while processing over 100 times faster (the latter is slowed down by heavy 3D volume processing). The B-variant outperforms MonoLSS, the strongest baseline without extra data, across almost all accuracy metrics while being 34% faster. Comparison with SOTA on the KITTI test set (largest X scale):

Method	Extra Data	Easy	Mod.	Hard
MonoTAKD	LiDAR	27.91	19.43	16.51
MonoDGP	Geometry	26.35	18.72	15.97
MonoDiff	—	30.18	21.02	18.16
LeAD-M3D X (Ours)	—	30.76	21.20	18.76

LeAD-M3D X outperforms all methods on AP3D|R40 (including those relying on LiDAR or geometric priors) while running 3.6 times faster than the previous SOTA method MonoDiff. On Roadside scenarios (Rope3D dataset), it achieves a top Car AP of 16.45. On the Waymo validation set, the X-variant achieves an AP0.5 (3D, L1) of 16.46, surpassing the previous best, MonoLSS, by 2.97. For cross-domain evaluation (nuScenes→KITTI), the B-variant scores an Easy AP0.5 of 45.50, outperforming MonoGDG (33.48), which is specifically designed for domain generalization, and even approaching unsupervised domain adaptation (UDA) methods that utilize target domain data.

Ablation Study¶

Component-wise ablation of CM3D and A2D2 (KITTI val set, B-variant, Car AP0.7 Mod. and Median Depth Error MDE):

Config	A2D2	CM3D-2D	CM3D-3D	Mod. AP↑	MDE↓(cm)
1 (Baseline YOLOv10-M3D B)		✓		19.60	61
2 (+CM3D-3D)		✓	✓	20.43	60
5 (Full LeAD-M3D B)	✓	✓	✓	22.65	56

Detailed ablation of A2D2 (B-variant, Mod. AP):

Config	Mod. AP↑	Description
Full A2D2	22.65	—
Replace with backbone features	22.30	Depth features are more aligned distillation targets
w/o quality metric	22.28	Drops 0.37%
w/o importance metric	22.04	Drops 0.61%
Replace with online self-distillation	21.80	Drops 0.85%, strong targets only emerge late
Teacher also processes MixUp (w/o asymmetry)	21.57	Drops 1.08%, asymmetry/denoising is indeed effective
w/o A2D2	20.43	—

Key Findings¶

A2D2 serves as the core engine: adding A2D2 on top of Config 2 (which integrates CM3D) increases Moderate AP by 2.22% and reduces depth error by an additional 4 cm. This makes it the primary driver of depth and accuracy improvements. CM3D independently contributes as well (Config 2 gains +0.83% AP and reduces MDE by 1 cm over the baseline).
Harder augmentation yields stronger distillation: MixUp creates the largest depth error gap for the teacher (75 cm vs. 66 cm for CutMix, 62 cm for RandAugment, and 59 cm for clean images). Distillation with MixUp achieves the highest AP, with all three augmentation types outperforming non-distilled baselines.
CGI3D provides "almost free" speedup: on the N-variant, it cuts down the runtime by about 2/3 and reduces FLOPs by ~75% with identical AP. Because the effective receptive field of the regression heads is exactly \(3\times3\), cropping a \(3\times3\) patch preserves all necessary context.
Mechanistically, denoising and distillation operate complementarily: MixUp denoising forces the student to rely on MixUp-invariant spatial cues (such as object structure) rather than noisy background features (as shown by saliency maps), while depth feature distillation provides an effective guiding signal for this denoising process.

Highlights & Insights¶

Asymmetry can be created "in reverse": while previous works focus on feeding LiDAR to the teacher, this paper guides the student with noisy MixUp images and requires it to reconstruct the teacher's clean features. This translates distillation into a denoising task, yielding a depth gap previously only achievable with privileged modalities using purely images. This paradigm can be extended to any distillation scenario where "the teacher has clean inputs, and the student needs to build robustness."
The two weighting terms in the weighted L1 loss are highly practical: using relative depth error as the quality weight prevents close-range gradients from dominating, while utilizing the absolute weights of the depth head as channel importance directly pinpoints "which channels decide depth," providing a more goal-oriented alignment than standard feature mapping.
CGI3D achieves strictly lossless "sparse inference" by precisely leveraging receptive field properties (a \(3\times3\) convolution followed by two \(1\times1\) convolutions equates to exactly a \(3\times3\) receptive field). Utilizing architectural properties to achieve cost-free acceleration is a highly valuable trick.
Using MGIoU instead of standard 3D IoU for matching bypasses the classic "gradient vanishing when bounding boxes do not overlap" issue during early training or for small objects, with a training overhead of less than 2%.

Limitations & Future Work¶

The authors focus solely on the fully supervised setting, and although they mention that the distillation process can be extended to semi-supervised settings to scale M3D, this remains unverified.
The teacher model is fixed to the largest X-variant for offline, frozen distillation, requiring a two-stage training process (34 hours for the teacher + up to 60 hours for the student). The overall training overhead is non-negligible, and online self-distillation was shown to yield inferior results, leaving a single-stage alternative currently absent.
The framework relies on the property of MixUp to preserve geometric quantities to construct a lossless denoising task. If augmentations that disrupt geometric consistency are used instead, the foundational premise of this framework is invalidated.
Main experiments focus on vehicle-mounted and roadside driving datasets (KITTI, Waymo, Rope3D, nuScenes). Although the paper claims strong generalizability, extreme viewpoint changes (such as large pitch angles in UAV views) are only mentioned in related work and lack formal evaluation in this paper.

vs. LiDAR distillation like MonoDistill / MonoTAKD: These methods rely on privileged LiDAR inputs for the teacher to create a student-teacher gap. In contrast, this paper creates this gap under a pure image setting using MixUp denoising, which simplifies the pipeline, allows shared teacher-student architectures, and outperforms these LiDAR-reliant methods on the KITTI test set.
vs. ADD / FD3D: These methods also avoid LiDAR but feed ground-truth object locations or object-level depth maps to the teacher (still following the "more information for the teacher" route). This paper takes the opposite direction by "giving the student harder inputs."
vs. geometric prior methods like MonoUNI / MonoDGP: These methods rely on perspective priors such as "2D height \(\propto 1/\text{depth}\)" to regularize depth, which only hold under standard vehicle-mounted views and fail under viewpoint shifts. This paper does not use any geometric priors and explicitly supports arbitrary SO(3) orientations, leading to superior robustness in cross-domain scenarios and roadside views (Rope3D).
vs. MonoLSS (strongest baseline without extra data): Under purely 3D box supervision, the proposed B-variant significantly outperforms MonoLSS in accuracy while running 34% faster, owing to A2D2 distillation and task-aligned matching for accuracy, combined with CGI3D for speed reduction.

Rating¶

Novelty: ⭐⭐⭐⭐ The inverse perspective of "feeding noise to the student to construct asymmetric distillation" is novel and highly practical. The combination of weighted loss and MGIoU matching is solid.
Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive coverage with evaluations across four datasets, cross-view experiments, domain generalization, component-wise ablations, and augmentation/receptive field analyses, accompanied by Pareto frontier plots.
Writing Quality: ⭐⭐⭐⭐ Clear structure. The contributions of individual designs are well-quantified via ablations, and equations are closely integrated with the figures.
Value: ⭐⭐⭐⭐⭐ Simultaneously achieves state-of-the-art accuracy and real-time inference (60+ FPS) without relying on LiDAR or geometric priors, delivering direct utility for the deployment of monocular 3D detection.