FD-CanKD: Frequency-Decoupled Cross-Attention Distillation as a Refinement Prior for Compact Object Detectors

YoungJae Cheong; Jhonghyun An

cs.CVAug 19, 2026

FD-CanKD: Frequency-Decoupled Cross-Attention Distillation as a Refinement Prior for Compact Object Detectors

Authors: YoungJae Cheong, Jhonghyun An

Organizations: Department of AI and Software, Gachon University, Seongnam, Republic of Korea

Abstract

Compact object detectors are suitable for resource-constrained visual perception, but their limited representation capacity creates an accuracy gap relative to large models. Conventional detector distillation often relies on prediction-level supervision or a single feature-alignment target, such as response, distribution, correlation, or frequency-domain matching. Frequency-Decoupled Cross-Attention Knowledge Distillation (FD-CanKD) is presented as a detector-oriented framework that transfers teacher knowledge at three complementary levels: head-level prediction supervision, relation-level non-local context transfer, and frequency-level component-selective alignment. Student features first aggregate teacher-side spatial context through cross-attention-based relation transfer, after which frequency-aware alignment preserves complementary structural and detail-sensitive cues. Under controlled Microsoft Common Objects in Context (COCO) experiments, fixed 50-epoch from-scratch comparisons show that FD-CanKD remains competitive with representative detector knowledge distillation baselines. Post-distillation continued fine-tuning further produces a stronger refinement-ready student than detector-only fine-tuning, reaching 48.87 mean average precision (mAP) at intersection-over-union thresholds from 0.50 to 0.95 (mAP50:95), 65.84 mAP50, and 53.40 mAP75 after 20 additional epochs. All distillation modules are removed after training, leaving the deployed student unchanged at 19.7M parameters. The framework is instantiated and evaluated in a controlled YOLOv12 teacher-student setting as a representative compact-detector case study.

Figures & tables

Figure 1: Comparison of knowledge distillation paradigms for YOLOv12 large-to-small distillation. (a) Head-only knowledge distillation (KD) provides prediction-level supervision only. (b) Direct feature KD performs point-wise feature matching between the teacher and student. (c) FD-CanKD combines head KD, relation-aware transfer, and band-specific feature alignment within a detector-oriented training pipeline 40 , 49 .

Figure 2: Overall architecture of the FD-CanKD integration framework. A frozen large-scale YOLOv12 teacher transfers knowledge to a compact YOLOv12 student 12 . The feature-level branch performs cross-attention-based non-local distillation on multi-scale detection features 40 , followed by fast Fourier transform (FFT)-based decomposition into low- and high-frequency components and mean squared error (MSE) and log-compressed MSE (LogMSE) alignment 49 . The prediction-level branch distills classification, box regression, and Distribution Focal Loss (DFL) outputs from the teacher head to the student head 58 .

Method	Feature KD Target	Head KD	$\mathbf{mAP}_{50:95}$	$\mathbf{mAP}_{50}$	$\mathbf{mAP}_{75}$
YOLOv12-M Student (50-epoch FS)	–	No	45.6	62.0	50.8
CWD + Head KD 30	Channel-wise distribution	Yes	46.3	62.4	50.5
MGD + Head KD 31	Masked generative feature	Yes	46.3	62.3	50.4
PKD + Head KD 24	Pearson correlation	Yes	46.2	62.2	50.5
FGD + Head KD 29	Focal/global feature	Yes	46.3	62.2	50.4
FreeKD + Head KD 36	Frequency-domain feature	Yes	46.5	62.7	50.7

Table 1: Controlled comparison with the locally reproduced 50-epoch from-scratch YOLOv12-M student reference and representative detector KD methods under the same YOLOv12-X to YOLOv12-M local training protocol. External feature-distillation methods are combined with the same head-output distillation branch for a fair controlled comparison.

Method	Rel.	Head	Freq.	$\mathbf{mAP}_{50:95}$	$\mathbf{mAP}_{50}$	$\mathbf{mAP}_{75}$
YOLOv12-M Student (50-epoch FS)	–	–	–	45.6	62.0	50.8
CanKD 40	Yes	No	No	46.5	62.4	50.5
CanKD + Head KD	Yes	Yes	No	46.3	62.4	50.5
FD-CanKD	Yes	Yes	Yes	46.5	62.8	50.8

Table 2: Selected internal component comparison under the controlled 50-epoch from-scratch YOLOv12-X to YOLOv12-M local training protocol. The table separates relation-aware feature transfer, head-output distillation, and frequency-aware alignment.

Method	Low	High	Low Loss	High Loss	$\mathbf{mAP}_{50:95}$	$\mathbf{mAP}_{50}$	$\mathbf{mAP}_{75}$
YOLOv12-M Student (50-epoch FS)	–	–	–	–	45.6	62.0	50.8
Rel. KD + Head	–	–	–	–	46.3	62.4	50.5
All-MSE	Yes	Yes	MSE	MSE	46.5	62.4	50.6
Low-only	Yes	No	MSE	–	46.5	62.6	50.7
High-only	No	Yes	–	LogMSE	46.4	62.5	50.3
Low + High-log	Yes	Yes	MSE	LogMSE	46.5	62.4	50.8

Table 3: Frequency-decoupled alignment ablation after relation-aware feature transfer under the controlled 50-epoch from-scratch teacher–student setting.

Scale Pair	Teacher $\mathbf{mAP}_{50:95}$	Student Base $\mathbf{mAP}_{50:95}$	FD-CanKD $\mathbf{mAP}_{50:95}$
YOLOv12-L $\rightarrow$ YOLOv12-M	47.0	45.6	46.4
YOLOv12-X $\rightarrow$ YOLOv12-M	49.5	45.6	46.5
YOLOv12-X $\rightarrow$ YOLOv12-S	49.5	40.9	41.1

Table 4: Scale-pair sensitivity analysis across different YOLOv12 teacher–student pairs under the controlled 50-epoch from-scratch local protocol.

Method	Init	Stage / Epochs	$\mathbf{mAP}_{50:95}$	$\mathbf{mAP}_{50}$	$\mathbf{mAP}_{75}$
YOLOv12-M FT	–	+5 FT	46.66	63.15	50.91
		+10 FT	46.71	63.25	50.81
		+20 FT	46.90	63.50	51.01
		+30 FT	47.02	63.68	51.20
CanKD	WS	Distilled, before FT	48.41 $\pm$ 0.12	65.09 $\pm$ 0.13	52.79 $\pm$ 0.18
		+5 FT	48.50	65.23	52.76

Table 5: Post-distillation continued fine-tuning analysis under the controlled local protocol. The table compares detector-only continued fine-tuning with warm-start CanKD and FD-CanKD trajectories, together with locally evaluated student and teacher references. The distilled before-FT rows report the mean and sample standard deviation over three random seeds.

Method	Stage	Setting	$\mathbf{mAP}_{50:95}$	$\mathbf{AP}_{S}$	$\mathbf{AP}_{M}$	$\mathbf{AP}_{L}$
YOLOv12-M Student (ref.)	Reference	Local student	45.63	28.92	51.94	62.35
CanKD	+20 FT	Relation-aware	48.75	29.19	53.76	66.09
FD-CanKD	+20 FT	Relation + frequency	48.87	30.15	53.54	65.41
CanKD	+30 FT	Relation-aware	48.80	29.44	53.87	66.23
FD-CanKD	+30 FT	Relation + frequency	48.84	30.02	53.43	65.43

Table 6: Object-size-wise COCO AP for the local YOLOv12-M student reference and the +20- and +30-epoch continued fine-tuning checkpoints.

Figure 3: Qualitative comparison on cluttered and occluded scenes. From left to right in each group, the results correspond to the YOLOv12-M student, the relation-aware KD variant, and FD-CanKD. The baseline student often shows lower confidence or unstable predictions in cluttered regions, while relation-aware KD improves several object responses. FD-CanKD combines relation-aware feature transfer with band-specific alignment and shows more stable detections, including fine-grained local object responses, in the illustrated examples.

Figure 4: Qualitative comparison on challenging scale and illumination variations. From left to right in each group, the results correspond to the YOLOv12-M student, the relation-aware KD variant, and FD-CanKD. The relation-aware KD variant improves teacher–student spatial transfer, while FD-CanKD additionally applies different alignment rules to decomposed feature components. The examples illustrate more balanced detections under scale and illumination changes, especially for local objects that require fine spatial detail.

Figure 5: Qualitative comparison on large objects and context-sensitive objects. From left to right in each group, the results correspond to the YOLOv12-M student, the relation-aware KD variant, and FD-CanKD. The proposed method maintains reliable detection of dominant objects while improving the stability of context-sensitive and fine-grained object predictions in the illustrated examples.

Explore similar work

FreqKD: Frequency-Decoupled Cross-Modal Knowledge Distillation for Infrared Object Detection
Jun 10, 2026Keval Thaker, Venkatraman Narayanan, Abdalmalek Aburaddaha +1Cross-Modal Knowledge Distillation Multispectral Object Detection
University of Michigan-Dearborn 4901 Evergreen Rd Dearborn, MI 48128, USA
OD3: Optimization-free Dataset Distillation for Object Detection
Jun 2, 2025Salwa K. Al Khatib, Ahmed ElHagry, Shitong Shao +1Dataset Distillation
Mohamed bin Zayed University of Artificial Intelligence · Hong Kong University of Science and Technology (Guangzhou)
CSCWD: Cross-Scale Channel-wise Knowledge Distillation for Lightweight Tiny Object Detection on Edge Devices
Sep 24, 2026Amir Zamani, Zeinab Ghasemi-NaraghiSmall Object Detection Efficient Object Detection
Department of Computer Engineering, Islamic Revolution Comprehensive University, Tehran, Iran · Department of Computer Engineering, K.N. Toosi University of Technology, Tehran, Iran

Jun 10, 2026, cs.CVJ/K move · Enter open · S save

FreqKD: Frequency-Decoupled Cross-Modal Knowledge Distillation for Infrared Object Detection

Keval Thaker, Venkatraman Narayanan, Abdalmalek Aburaddaha, Samir A. Rawashdeh

University of Michigan-Dearborn 4901 Evergreen Rd Dearborn, MI 48128, USA

Cross-Modal Knowledge Distillation Multispectral Object Detection Frequency-Domain Feature Learning Infrared Object Detection Knowledge Distillation

Open paper AskPDF arXiv

Transfer learning from large-scale RGB foundation models to infrared (IR) imagery through knowledge distillation (KD) remains challenging due to fundamental differences in image formation physics. We investigate the spectral structure of the RGB--IR modality gap and observe that feature divergence is not uniform across spatial frequencies: low-frequency components (shape, layout) show greater cross-modal alignment than high-frequency components (texture, fine edges), which reflect modality-specific characteristics. Based on this analysis, we propose FreqKD, a frequency-decoupled distillation framework that applies asymmetric supervision adapted to each band's cross-modal consistency. The method employs strict mean squared error (MSE) on the low-frequency band to preserve shared structural information and a relaxed log-MSE loss (weighted at 0.1) on the high-frequency band to provide edge guidance while tolerating texture differences. Spectral divergence analysis on 500 paired samples shows that high-frequency divergence exceeds low-frequency divergence by a factor of 2.4x on average across all analysed transformer layers. On KAIST multispectral pedestrian detection, FreqKD achieves 64.1 mAP50, improving 2.4 points over the DINOv2 baseline. The learned representation transfers across datasets (FLIR ADAS, +2.1 mAP50), tasks (MFNet segmentation, +1.85 mean intersection-over-union), and architectures (ResNet-50, +1.0 mAP50). Code is available at: https://anonymous.4open.science/r/freq_decoupled_kd-5E5A

Abstract

Explore similar work

FreqKD: Frequency-Decoupled Cross-Modal Knowledge Distillation for Infrared Object Detection

OD3: Optimization-free Dataset Distillation for Object Detection

CSCWD: Cross-Scale Channel-wise Knowledge Distillation for Lightweight Tiny Object Detection on Edge Devices