Research map / Label-efficient and robust learning / Domain adaptation and generalization
EAGLE: Efficient Adaptive Geometry-based Learning in Cross-view Understanding
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 EAGLE, an unsupervised cross-view adaptation method for open-vocabulary semantic segmentation (car-view to drone-view) using a geodesic-flow-based cross-view geometric constraint on unpaired data plus view-condition prompts.
Problem
Semantic segmentation and open-vocabulary segmentation models fail to generalize across camera views (car/ground view to UAV/drone view) because existing UDA methods lack cross-view geometric structure modeling.
Approach
Impose a cross-view geometric constraint between images and segmentation masks on unpaired source/target samples, measure structural change with a geodesic-flow-based correlation metric on Grassmann subspaces, and add view-condition text prompts ("captured from the [domain] view") whose embedding correlations are constrained by the same geometry.
Key results
- On SYNTHIA to UAVID with Mask2Former, EAGLE with view-condition prompting reaches 52.2% mIoU, versus 13.2% without adaptation and 20.2% for prompted no-adaptation.
- With DAFormer on SYNTHIA to UAVID, EAGLE reaches 50.8% mIoU, outperforming CROVIA (46.0%) and MIC (41.8%).
- On BDD to UAVID, EAGLE with Mask2Former-style open-vocab segmentation reaches 50.7% mIoU versus 47.9% for FreeSeg + Cross-View.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add cross-view adaptation to prompting (ablation) | Mask2Former segmentation · SYNTHIA → UAVID | mIoU: 20.2 -> 51.1 | Helped |
| Add cross-view adaptation to prompting (ablation) | Mask2Former segmentation · GTA → UAVID | mIoU: 18.1 -> 44.0 | Helped |
| Add view-condition prompting (ablation) | EAGLE with cross-view adaptation · SYNTHIA → UAVID | mIoU: +1.1% | Helped |
| Add view-condition prompting (ablation) | EAGLE with cross-view adaptation · GTA → UAVID | mIoU: +1.7% | Helped |
| Use Geodesic Flow metric instead of Euclidean metric (ablation) | EAGLE · SYNTHIA → UAVID | mIoU: 27.2 -> 52.2 | Helped |
| Use Geodesic Flow metric instead of Euclidean metric (ablation) | EAGLE · GTA → UAVID | mIoU: 26.5 -> 45.7 | Helped |
| Use a Swin backbone instead of ResNet (ablation) | EAGLE with Geodesic Flow metric · SYNTHIA → UAVID | mIoU: 52.2 -> 54.7 | Helped |
| Use a Swin backbone instead of ResNet (ablation) | EAGLE with Geodesic Flow metric · GTA → UAVID | mIoU: 45.7 -> 47.1 | Helped |
| Set α to 1.5 rather than 0.5, 1.0, or 2.0 (ablation) | EAGLE · SYNTHIA → UAVID | mIoU: 49.4; 51.4; 52.2; 50.6 | Helped |
| Set α to 1.5 rather than 0.5, 1.0, or 2.0 (ablation) | EAGLE · GTA → UAVID | mIoU: 41.1; 42.4; 45.7; 44.2 | Helped |
| Increase subspace dimension from 96 to 256 (ablation) | EAGLE · SYNTHIA → UAVID | mIoU: 49.6 -> 52.2 | Helped |
| Increase subspace dimension from 96 to 256 (ablation) | EAGLE · GTA → UAVID | mIoU: 42.3 -> 45.7 | Helped |
Surprising findings
- Euclidean-metric variants and geodesic-flow variants differ markedly, with geodesic flow giving roughly +25 mIoU over no-adaptation in some settings.
- Batch size strongly affects the subspace estimation, with batch size 4 giving 35.2% mIoU versus 52.2% at batch size 16.
Limitations
- Relies on a linear scale hypothesis between image and segmentation structural changes (alpha, gamma), which is not theoretically verified.
- Limited diversity of classes in the benchmarks (only five to six shared classes).
- Hyper-parameters such as loss weights and subspace dimension are not fully explored; batch size limited by GPU resources.
Experimental setup
Datasets: SYNTHIA GTA BDD100K UAVID
Benchmarks: SYNTHIA to UAVID GTA to UAVID BDD to UAVID
Metrics: mIoU per-class IoU
Baselines: AdvEnt Polar Transform DADA BiMaL SAC ProDA CROVIA DAFormer MIC DenseCLIP FreeSeg
Builds on: Mask2Former FreeSeg CLIP DenseCLIP geodesic flow kernel
Closest papers by meaning
- UniDSeg: Unified Cross-Domain 3D Semantic Segmentation via Visual Foundation Models Prior (NeurIPS 2024 · similarity 0.88)
- Cross-Modal and Uncertainty-Aware Agglomeration for Open-Vocabulary 3D Scene Understanding (CVPR 2025 · similarity 0.87)
- Building a Strong Pre-Training Baseline for Universal 3D Large-Scale Perception (CVPR 2024 · similarity 0.86)
- Open-Set Domain Adaptation for Semantic Segmentation (CVPR 2024 · similarity 0.86)
- A Unified Framework for 3D Scene Understanding (NeurIPS 2024 · similarity 0.86)
- Large Spatial Model: End-to-end Unposed Images to Semantic 3D (NeurIPS 2024 · similarity 0.86)
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.