IMKD: Intensity-Aware Multi-Level Knowledge Distillation for Camera–Radar Fusion

Shashank Mishra · Karan Patil · Didier Stricker · Jason Rambach

WACV 2026

Project Page | Paper | Models

---

IMKD introduces a three-stage, intensity-aware multi-level knowledge distillation pipeline for high-performance Radar–Camera 3D object detection without requiring LiDAR at inference time.

---

🚀 Abstract

High-performance Radar–Camera 3D object detection can be achieved through knowledge distillation without using LiDAR at inference time. However, existing distillation methods often transfer modality-specific features directly to each sensor, potentially distorting their unique characteristics and diminishing their strengths.

We propose IMKD, a radar–camera fusion framework based on multi-level, intensity-aware knowledge distillation that maintains each sensor’s intrinsic properties while enhancing their complementarity.

IMKD introduces a three-stage distillation strategy:

1. LiDAR → Radar intensity-aware feature distillation
   Enhances radar representations with detailed structural cues.

2. LiDAR → Fused intensity-guided distillation
   Selectively emphasizes valuable geometric and depth information at the fusion layer, encouraging complementary sensor interaction rather than strict alignment.

3. Camera–Radar intensity-guided fusion mechanism
   Facilitates improved calibration and alignment between modalities.

Experiments on the nuScenes benchmark show that IMKD achieves:

• 67.0% NDS
• 61.0% mAP

surpassing all previous distillation-based radar–camera fusion approaches.

Our code and models will be publicly released.

---

🏗️ Architecture

---

Knowledge Distillation Performance on nuScenes

This section compares existing knowledge distillation (KD) methods for 3D object detection on the nuScenes test set. We focus on the impact of different distillation strategies while using each method’s best reported backbone and input resolution.

Method	Input	KD	NDS ↑	mAP ↑	mATE ↓	mASE ↓	mAOE ↓	mAVE ↓	mAAE ↓
UVTR (Li et al. 2022)	C	L2C	52.2	45.2	0.612	0.256	0.385	0.664	0.125
BEVDistill (Chen et al. 2022)	C	LC2C	59.4	49.8	0.472	0.247	0.378	0.326	0.125
UniDistill (Zhou et al. 2023)	C	L2C	39.3	29.6	0.637	0.257	0.492	1.084	0.167
LabelDistill (Kim et al. 2024)	C	LL◇2C	61.0	52.6	0.443	0.252	0.339	0.370	0.136
X3KD (Klingner et al. 2023)	C	LC2C	56.1	45.6	0.506	0.253	0.414	0.366	0.131
X3KD (Klingner et al. 2023)	C+R	L2CR	55.3	44.1	–	–	–	–	–
CRKD (Zhao et al. 2024)	C+R	LC2CR	58.7	48.7	0.404	0.253	0.425	0.376	0.111
IMKD (Ours)	C+R	LL◇2M	67.0	61.0	0.401	0.249	0.305	0.238	0.102

Notes:
L = LiDAR, L◇ = Label, C = Camera, R = Radar, M = Merged (Camera + Radar).
Arrows indicate whether higher (↑) or lower (↓) values are better.

3D Object Detection Performance on nuScenes (Validation Set)

This section compares camera- and radar-based 3D object detection methods on the nuScenes validation set under fair evaluation settings. All methods exclude future frames, test-time augmentation, and CBGS to ensure a consistent comparison. Results are grouped by backbone and image resolution.

ResNet-50 Backbone (BEVDepth-compatible setting)

Method	Input	Backbone	Image Size	NDS ↑	mAP ↑	mATE ↓	mASE ↓	mAOE ↓	mAVE ↓	mAAE ↓
BEVDet (Huang et al. 2021)	C	ResNet-50	256 × 704	39.2	31.2	0.691	0.272	0.523	0.909	0.247
BEVDepth (Li et al. 2023)	C	ResNet-50	256 × 704	47.5	35.1	0.639	0.267	0.479	0.428	0.198
RC-BEVFusion (Stacker et al. 2023)	C+R	ResNet-50	256 × 704	52.5	43.4	0.511	0.270	0.527	0.421	0.182
SOLOFusion (Park et al. 2022)	C	ResNet-50	256 × 704	53.4	42.7	0.567	0.274	0.411	0.252	0.188
StreamPETR (Wang et al. 2023)	C	ResNet-50	256 × 704	54.0	43.2	0.581	0.272	0.413	0.295	0.195
SparseBEV (Liu et al. 2023)	C	ResNet-50	256 × 704	54.5	43.2	0.606	0.274	0.387	0.251	0.186
CRN (Kim et al. 2023)	C+R	ResNet-50	256 × 704	56.0	49.0	0.487	0.277	0.542	0.344	0.197
RCBEVDet (Lin et al. 2024)	C+R	ResNet-50	256 × 704	56.8	45.3	0.486	0.285	0.404	0.220	0.192
CRT-Fusion (Kim et al. 2024)	C+R	ResNet-50	256 × 704	57.2	50.0	0.499	0.277	0.531	0.261	0.192
IMKD (Ours)	C+R	ResNet-50	256 × 704	61.0	51.6	0.444	0.259	0.384	0.229	0.160
RICCARDO (Long et al. 2025)	C+R	ResNet-101	1408 × 512	62.2	54.4	0.481	0.266	0.325	0.237	0.189
IMKD (Ours)	C+R	ResNet-101	1408 × 512	62.7	53.9	0.417	0.255	0.348	0.235	0.158

Notes:

• C = Camera, R = Radar
• ↑ / ↓ indicate higher / lower is better
• Upper block: comparisons restricted to BEVDepth-style ResNet-50 settings
• Lower block: ResNet-101 with higher image resolution, included for completeness

🔧 Getting Started

Code and models will be made public soon.

Contact

You can contact the author through email: Shashank.Mishra@dfki.de

Citing

If you find our work useful, please consider citing:

@misc{mishra2025imkdintensityawaremultilevelknowledge,
      title={IMKD: Intensity-Aware Multi-Level Knowledge Distillation for Camera-Radar Fusion}, 
      author={Shashank Mishra and Karan Patil and Didier Stricker and Jason Rambach},
      year={2025},
      eprint={2512.15581},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.15581}, 
}