Socialized Detector Learning: Trajectory-Guided and Reciprocal Distillation for Heterogeneous Object Detectors

Weihao Li; Yunqi Zhu; Zhihe Fan; Ruipu Zhao; Boan Tao; Xinjie Yao; Yan Fan; Pengfei Zhu

cs.CVAug 26, 2026

Socialized Detector Learning: Trajectory-Guided and Reciprocal Distillation for Heterogeneous Object Detectors

Authors: Weihao Li, Yunqi Zhu, Zhihe Fan, Ruipu Zhao, Boan Tao, Xinjie Yao, Yan Fan, Pengfei Zhu

Organizations: School of New Media and Communication, Tianjin University, Tianjin, China · School of Computer Science and Engineering, University of New South Wales, Sydney, Australia · School of Sports Training, Tianjin University of Sport, Tianjin, China · School of Artificial Intelligence, Tianjin University, Tianjin, China · School of Automation, Southeast University, Nanjing, China · School of Computer Science and Technology, Tianjin University, Tianjin, China · Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming, China · School of Electronic Science, National University of Defense Technology, Hunan, China

Abstract

Object detection knowledge is fragmented across independently trained, heterogeneous detectors with complementary category supports. In socialized learning, this knowledge resides in a society, and learning aims to evolve the society collectively through exchange. However, aggregation-based socialization does not explicitly plan transfer order, whereas progressive multi-teacher distillation considers order but remains a one-way student enhancement in a shared category space. Building on Socialized Learning, we formulate Socialized Detector Learning (SDL) for heterogeneous, category-specialized object detectors and propose Trajectory-Guided and Reciprocal Distillation (TGRD).TGRD estimates directed operational Inter-Detector Transfer Difficulty (IDTD) from held-out feature-alignment residuals, precomputes a fixed score table, and greedily constructs a carrier trajectory. Along the trajectory, knowledge is progressively consolidated into a union-category carrier and then returned to experts through reciprocal transfer. A conditional proxy-certificate analysis shows that, under stated assumptions, the progressive certificate is no larger than an aggregated-target counterpart. On MS COCO with four heterogeneous experts and two carrier initializations, final carriers outperform epoch-matched simultaneous aggregation controls by 2.6 AP in both settings. Reciprocal detectors attain 20.8--28.4 AP on previously unsupported categories while remaining within 1.3 AP of original expert-specific performance. These results support order-aware progressive consolidation followed by reciprocal transfer as a viable mechanism for detector-society evolution.

Figures & tables

Figure 1: Detector-society evolution under SDL. TGRD progressively consolidates complementary expertise and reciprocally transfers it back, broadening category coverage while preserving specialization.

Figure 2: Compatibility-guided carrier trajectory planning. Held-out alignment residuals define a fixed table of directed operational scores $\widehat{D}(A,B)$ over ordered anchor–candidate pairs, from which greedy selection determines the expert order and carrier trajectory.

Figure 3: Progressive and reciprocal knowledge transfer in TGRD. Left: Along the planned trajectory, experts transfer knowledge via AFD to a carrier whose prediction head expands over accumulated categories, yielding $S_{K}$ . Right: A reciprocal detector initialized from each expert is guided by the corresponding frozen original expert and the frozen $S_{K}$ ; the updated detectors form $\mathcal{T}^{(r+1)}$ .

(a) Operational IDTD scores $\widehat{D}(a_{k-1},t_{i})$
	Candidate next expert $t_{i}$
Planning anchor $a_{k-1}$	$t_{1}$	$t_{2}$	$t_{3}$	$t_{4}$
$S_{0}$ (RetinaNet)	0.5463	1.2115	0.5588	0.6411
$S_{0}$ (Faster R-CNN)	0.5148	1.1658	0.2318	0.5473
$t_{1}$ (RetinaNet)	–	1.1792	0.5514	0.4862
$t_{2}$ (FCOS)	1.1392	–	1.0478	1.1748

Table 1: Operational IDTD scores and fixed-table greedy trajectories under two carrier initializations. Detector architectures are given in parentheses; all configurations use R50–FPN.

(a) RetinaNet as initial carrier $S_{0}$
Carrier state / control	$AP$	$AP_{\mathcal{C}_{0}}$	$AP_{\Delta\kappa_{1}}$	$AP_{\Delta\kappa_{2}}$	$AP_{\Delta\kappa_{3}}$	$AP_{\Delta\kappa_{4}}$
$S_{0}$	37.3	37.3	–	–	–	–
$S_{1}\,[t_{1}]$	35.9	39.2	22.8	–	–	–
$S_{2}\,[t_{1},t_{4}]$	35.1	40.5	22.6	26.4	–	–
$S_{3}\,[t_{1},t_{4},t_{3}]$	34.5	39.6	19.8	39.4	24.1	–
$S_{4}\,[t_{1},t_{4},t_{3},t_{2}]$	32.8	39.5	21.0	22.2	38.7	22.0

Table 2: Stage-wise carrier performance from

S_{0}

along two fixed-table trajectories versus simultaneous aggregation (Avg-FPN KD). Overall AP is evaluated on

\mathcal{K}_{k}

for

S_{k}

and on

\mathcal{C}_{\cup}

for the control, with

\mathcal{K}_{0}=\mathcal{C}_{0}

AP_{\Delta\kappa_{j}}

denotes AP over the categories introduced by

t_{\pi(j)}

that are absent from

\mathcal{K}_{j-1}

; control columns follow the same trajectory-specific order. Brackets list the incorporated experts.

	Initial support $\mathcal{C}_{i}$			Expert-specific support $\mathcal{C}_{i}\setminus\mathcal{C}_{0}$			Knowledge expansion
Expert	Original	Reciprocal	$\Delta$	Original	Reciprocal	$\Delta$	$AP_{\mathcal{C}_{\cup}\setminus\mathcal{C}_{i}}$	$AP_{\mathcal{C}_{\cup}}$
(a) Final RetinaNet carrier $S_{4}$
$t_{1}$	35.5	35.9	$+0.4$	22.5	22.3	$-0.2$	25.8	32.1
$t_{2}$	40.0	39.2	$-0.8$	31.7	30.6	$-1.1$	25.6	34.1
$t_{3}$	39.8	40.1	$+0.3$	40.2	40.7	$+0.5$	23.2	33.7
$t_{4}$	38.8	38.9	$+0.1$	26.5	27.8	$+1.3$	28.4	35.0

Table 3: Expert retention and knowledge expansion after reciprocal updates jointly guided by the original experts and final (a) RetinaNet or (b) Faster R-CNN carriers.

\Delta

denotes the reciprocal-minus-original AP change.

Explore similar work

Jul 15, 2026cs.CV

Symbiosis-Inspired Knowledge Distillation for Incremental Object Detection

Incremental object detection (IOD) aims to extend detectors to new categories while retaining previously acquired knowledge. Existing methods often adopt a class incremental learning perspective, separating feature spaces to sharpen decision boundaries. However, this separation-oriented paradigm may overlook object symbiosis in detection, where co-occurrence and occlusion introduce spatial and semantic dependencies that benefit from shared representations. Ignoring these dependencies distorts the shared representations, exacerbates confusion between old and new classes, and accelerates catastrophic forgetting. To address this, we propose Symbiosis-Inspired Knowledge Distillation (SIKD), which explicitly leverages object symbiosis at two complementary levels. Spatial Symbiosis Distillation (SpSD) focuses on symbiotic regions where the old model responds with high overlap to objects in the new task. It preserves generalizable old class cues, suppresses class-specific bias and redundancy, and distills the refined evidence to the new model at matched spatial locations with slot-aligned supervision. Semantic Symbiosis Distillation (SeSD) maintains class level structure by forming confidence weighted prototypes for old classes and aligning their inter class soft ranks over the old class logits, which stabilizes the semantic topology during adaptation. Extensive experiments demonstrate the effectiveness and superiority of the proposed method.

Mingyue Zeng, De Cheng, Zhipeng Xu +3

State Key Laboratory of Integrated Services Networks, School of Telecommunications Engineering, Xidian University, Xi’an, China · School of Electronic Engineering, Xidian University, Xi’an, China

Class-Incremental Learning Object Detection Knowledge Distillation Catastrophic Forgetting

Ask PDF

Sep 17, 2026cs.CV

Enhanced Knowledge Distillation for Detection Transformer via Teacher Prediction Refinement

Detection Transformers (DETRs) achieve strong performance in object detection but remain challenging to deploy on edge devices due to their high computational cost. Existing DETR distillation methods mainly focus on aligning distillation points, while largely overlooking the quality of the teacher's supervision itself. We observe that due to stage-wise non-monotonic prediction behavior in DETRs, well-localized or correctly classified predictions from earlier stages may degrade in later ones, and some negative predictions become increasingly overconfident. As a result, relying solely on the current stage's predictions yields inaccurate and inconsistent supervision. To address this issue, we propose Teacher Prediction Refinement Distillation (TPRD), a plug-and-play module that refines teacher predictions before distillation by exploiting stage-wise prediction information. TPRD improves supervision quality through Positive Prediction Correction (PPC), which corrects degraded positive predictions by restoring more accurate ones from earlier stages, ensuring reliable localization and classification signals, and Negative Prediction Suppression (NPS) suppresses the influence of overconfident negatives, preventing them from providing misleading supervision to the student. To preserve informative dark knowledge, we further introduce Maximum Dark Knowledge Preservation (MDKP), which selectively refines target-class logits while retaining non-target relations. Extensive experiments on MS COCO and PASCAL VOC demonstrate the effectiveness and robustness of the proposed method. Our code is available at https://github.com/xingyitong1/TPRD.

Yitong Xing, Yuhao Cheng, Yanping Li +1

School of Computer Science, Shanghai Jiao Tong University, Shanghai 200240, China

Detection Transformer Object Detection Knowledge Distillation

Ask PDF

Aug 19, 2026cs.CV

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

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.

YoungJae Cheong, Jhonghyun An

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

Knowledge Distillation Efficient Object Detection Feature-Based Knowledge Distillation Relational Knowledge Distillation

Ask PDF