Skip to content
Book Open access

CT-CLIP: Calibrating Semantic Occupancy Tokens for Class-Guided Camouflaged Object Detection

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 6 references

TL;DR

CT-CLIP is proposed, a novel framework for class-guided camouflaged object detection that calibrates SOT to restore spatial structure and significantly outperforms existing methods on the CGCOD benchmarks.

Abstract

Class-Guided Camouflaged Object Detection (CGCOD) leverages category-level semantic priors to improve model perception in complex camouflage scenes. However, when applying vision-language models like CLIP to pixel-level segmentation, the deep layers of its Vision Transformer generate numerous Semantic Occupancy Tokens (SOT), degrading the spatial structure of visual features. This further exacerbates the optimization conflict between semantic discrimination and pixel-wise localization, becoming a critical bottleneck that limits fine-grained segmentation performance. To address this, we propose CT-CLIP, a novel framework for class-guided camouflaged object detection that calibrates SOT to restore spatial structure. At its core lies the Multi-stage Attention Calibration Module, which sequentially mitigates SOT-induced spatial degradation through three synergistic operations: SOT Reallocation redistributes attention from SOT to object regions; Hierarchical Attention Fusion adaptively aggregates rectified multi-layer attention maps to enhance last-layer spatial perception; and DINO-Guided Augmentation injects self-supervised attention from DINO as fine-grained structural priors, achieving high-fidelity attention reconstruction. Building on the restored features, a Semantic Query Decoder is devised to reformulate segmentation as a semantic-guided visual querying process, where text embeddings serve as queries, and the concatenation of CLIP visual features with DINO-enhanced features serves as keys and values in cross-attention, thereby progressively aligning semantics with spatial structures. Experimental results on the CGCOD benchmarks demonstrate that CT-CLIP significantly outperforms existing methods.

Read PDF

Similar papers

Open access Sep 2026

Vision–language guided semantic-geometric transformer for memory-efficient 3D scene understanding

This work proposes a segmentation framework guided by Contrastive Language–Image Pre-training (CLIP) that enriches sparse 3D tokens with vision–language semantic priors and introduces a decoupled CLIP-induced semantic residual that forms semantic-geometric attention biases for local window attention.

Li-Cheng Liu, Yu Li, Fu-Yong Liu · 0 citations
Open access Nov 2026

Lightweight Architecture Optimization for Car Logo Detection Based on Visual State Space Models

Car logo detection is a challenging fine-grained perception task characterized by small object scales and high inter-class similarity. In such scenarios, discriminative cues mainly reside in subtle local textures rather than global structures. Although State Space Model (SSM)-based detectors, such as Mamba-YOLO with it...

Jing-Wen Gong, Jun-Xing Zhang, Chun-Juan Bo · 0 citations
Open access Oct 2026

HyMTAS-GBM: Integrating Semantic Convolutions and Multi-Head Attention with Gradient Boosting for Robust Object Detection

Object detection in complex scenes requires models that can maintain precise spatial attention and reliable query representations. However, existing CNN and DETR-based detectors often exhibit diffuse cross-attention, weak query initialization, and limited robustness in cluttered or dynamic environments. This paper prop...

Rahul Mandlik, Avinash Sharma · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.