cs.CVAug 22, 2026

C2^2Path: Class-Conditional Pathway Decoupling for Vision-Language Incremental Object Detection

Authors: Lecheng Xu, Feifei Shao, Ouyangzi Ye, Zhen Wang, Lin Li, Kexin Li, Zhao Wang, Changqin Huang

Organizations: Zhejiang University · The Hong Kong University of Science and Technology · Zhejiang Tobacco Monopoly Administration

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Incremental Object Detection (IOD) aims to enable detectors to continuously learn novel categories while preserving previously acquired knowledge. However, existing methods suffer from two forms of class knowledge coupling: class boundary erosion induced by shared parameter updates and class representation entanglement arising from mixed feature encoding. We argue that effective incremental learning requires class-specific computational pathways that enable isolated parameter updates and separated class-wise injection. To this end, we propose \textbf{C2^2Path}, a class-conditional pathway decoupling framework for vision-language incremental object detection that leverages token-level class cues to establish dedicated and updatable computational pathways for different categories. Specifically, C2^2Path introduces a category expert library and a class-conditional decoupling module. The expert library consists of learnable low-rank computational nodes that capture category-specific knowledge, while the decoupling module generates class-aware routing signals to dynamically compose ClassLoRA adapters from these experts, thereby forming class-specific computational pathways for isolated updates and separated injection across categories. Extensive experiments on COCO 2017 under multiple incremental learning settings demonstrate that C2^2Path consistently outperforms state-of-the-art methods, providing an effective and scalable solution for continual category expansion in vision-language detectors.

Figures & tables

Figure 1: Illustration of class knowledge coupling problem and class-conditional pathway decoupling solution. Existing methods suffer from class boundary erosion and representation entanglement. C 2 Path dynamically selects category-specific experts through class-conditional routing to construct independent pathways, enabling novel-category adaptation while preserving existing class knowledge.
Figure 2: Overview of the C 2 Path framework. Left: the overall incremental learning framework with multi-level knowledge distillation. Right: the proposed category-conditioned parameter synthesis mechanism, where the Class-Conditional Decoupling Module ( CCDM ) transforms intermediate text features into token-conditioned activation signals, which dynamically combine low-rank experts in the Class Expert Library ( CEL ) to synthesize token-specific ClassLoRA instances.
SettingMethodDetectorAPAPAP50AP_{50}AP75AP_{75}APSAP_{S}APMAP_{M}APLAP_{L}
40+40CL-DETR 21Deformable DETR42.060.145.924.045.355.6
DMD 11Deformable DETR39.8-----
SDDGR 13Deformable DETR43.062.147.124.946.957.0
TALIR 35GLIP ∗40.457.443.923.344.754.5
LwF 17Grounding DINO ∗21.829.423.712.423.030.4
ERD 4Grounding DINO ∗29.741.832.119.033.138.8
Table 1: Comparison with state-of-the-art methods on MS COCO 2017. ∗ denotes vision-language detectors.
Figure 3: The labels (a–d) show enlarged views of the object categories predicted in the four columns from left to right.
BaselineCELCCDMAPAPAP50AP_{50}AP75AP_{75}
GCD  46.763.950.8
√\surd 48.465.352.6
√\surd√\surd48.765.653.1
Table 2: Ablation study on MS COCO 2017 under the 70+10 setting. CEL: Category Expert Library. CCDM: Category-Conditional Decoupling Module. “Only CEL” denotes that ClassLoRA is generated by averaging all experts in CEL.
MethodPhase 1Phase 2
BaseBaseNovelAll
C 2 Path (Ours)55.853.040.046.5
Table 3: Performance Analysis under the 40+40 Setting. Phase 1 reports the AP of the base model on the 40 base categories, whereas Phase 2 reports the AP of the final model on the base (1–40), novel (41–80), and all (1–80) categories.
Figure 4: Visualization of visual–linguistic features for 20 representative categories (10 base and 10 novel) after incremental training. Phase 1 learns initial representations from 70 base categories, while Phase 2 performs exemplar-free adaptation to 10 novel categories. Blue ellipses and red rectangles denote base and novel categories, respectively.
Figure 5: Pairwise cosine distance matrices of the encoder-fused text features of the 70 base categories. Phase 1 learns initial representations from 70 base categories, while Phase 2 performs exemplar-free adaptation to 10 novel categories.
More in this topic

Continual Object Detection

See all 13 papers

Explore similar work

CardsList

  1. Beyond Prompt Degradation: Prototype-guided Dual-pool Prompting for Incremental Object Detection

    Mar 2, 2026Yaoteng Zhang, Zhou Qing, Junyu Gao +1Object DetectionClass-Incremental Learning

    Northwestern Polytechnical University

  2. CL-CLIP: CLIP-Based Continual Learning Framework with Cost-Volume Category Decoupling for Object Detection

    Jun 5, 2026Zihan Liu, Yuguang Yang, Shengjie Su +5VLM AdaptationContinual Object Detection

    National College for Excellent Engineers, Beihang University · School of Electronic Information Engineering, Beihang University · AI Research, Qihoo 360 +5

  3. Symbiosis-Inspired Knowledge Distillation for Incremental Object Detection

    Jul 15, 2026Mingyue Zeng, De Cheng, Zhipeng Xu +3Class-Incremental LearningObject Detection

    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