Research map / Detection and segmentation / Open-set and few-shot detection
Prompt-Driven Dynamic Object-Centric Learning for Single Domain Generalization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a prompt-driven dynamic object-centric network for single-domain generalization, using CLIP text prompts to gate ResNet features via Slot-Attention fusion and spatial/channel masks, evaluated on PACS classification and urban-scene object detection.
Problem
Single-domain generalization fails because static networks cannot adapt to varying scene complexity across unseen target domains.
Approach
A prompt-based object-centric gating module fuses CLIP scene-prompt text embeddings with visual features via Slot-Attention to produce binary gating masks that dynamically select spatial and channel features in a ResNet backbone, regularized by a bound loss on mask density.
Key results
- Single-domain PACS with ResNet-18: 71.12% average accuracy, versus 69.86% for Meta-Casual (prior best).
- Multi-domain PACS leave-one-domain-out: 88.07% average, versus 87.00% for Meta-Casual.
- Diverse-Weather object detection (Faster-RCNN ResNet-101): 39.1% average mAP, versus 38.5% for CLIP-Gap.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add dynamic gating network (no prompt, no Slot-Attention) over static ResNet-18 base (ablation) | ResNet-18 on PACS · PACS single-DG | Avg accuracy (%): 54.97 -> 64.27 | Helped |
| replace Slot-Attention fusion with traditional attention in the gating module (removing Slot-Attention) (ablation) | prompt-based dynamic gating · PACS single-DG | Avg accuracy (%): 71.12 -> 68.94 (normal attention, which is with prompt) ; Slot without prompt 66.81 | Hurt |
| remove prompt guidance from Slot-Attention gating (dynamic network with slot attention only) (ablation) | prompt-based dynamic gating · PACS single-DG | Avg accuracy (%): 71.12 -> 66.81 | Hurt |
| add prompt-based Slot-Attention object-centric gating on top of dynamic network (ablation) | ResNet-18 on PACS · PACS single-DG | Avg accuracy (%): 64.27 -> 71.12 | Helped |
| prompt-based object-centric dynamic network vs prior SOTA | ResNet-18 on PACS · PACS single-DG | Avg accuracy (%): +1.2 over Meta-Casual (69.86 -> 71.12) | Helped |
| prompt-based dynamic network applied to multi-domain generalization | ResNet-18 on PACS · PACS multi-DG | Avg accuracy (%): 88.07 vs 87.00 (Meta-Casual) | Helped |
| prompt-based dynamic object-centric network on Diverse-Weather object detection | Faster-RCNN ResNet-101 · Diverse-Weather single-DG detection (average over target domains) | mAP (%): 39.1 vs 38.5 (CLIP-Gap) | Helped |
| prompt-based dynamic network on Night Rainy target domain | Faster-RCNN ResNet-101 · Daytime Clear to Night Rainy | mAP (%): 19.2 vs 18.7 (CLIP-Gap), +0.5 | Helped |
Limitations
- Gains on object detection are modest (about 0.6 mAP over CLIP-Gap on average).
- Gating design is demonstrated only on ResNet; generality to other architectures is not shown in the text.
Experimental setup
Datasets: PACS Diverse-Weather Dataset
Benchmarks: PACS Diverse-Weather Dataset
Metrics: classification accuracy mAP
Baselines: RSC RSC+ASR L2D P-RC Meta-Casual Faster-RCNN IBN-Net IterNorm SW ISW S-DGOD CLIP-Gap
Builds on: CLIP Slot-Attention Dynamic networks (gating) Verelst et al. bound loss CoCoOp
Closest papers by meaning
- Towards Single-Source Domain Generalized Object Detection via Causal Visual Prompts (NeurIPS 2025 · similarity 0.89)
- Object-Aware Domain Generalization for Object Detection (AAAI 2024 · similarity 0.86)
- G-NAS: Generalizable Neural Architecture Search for Single Domain Generalization Object Detection (AAAI 2024 · similarity 0.86)
- Style Evolving along Chain-of-Thought for Unknown-Domain Object Detection (CVPR 2025 · similarity 0.85)
- SPTNet: An Efficient Alternative Framework for Generalized Category Discovery with Spatial Prompt Tuning (ICLR 2024 · similarity 0.85)
- Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.