Awesome Referring Remote Sensing Image Segmentation

:loudspeaker: Call for Contribution

We actively welcome:

• :page_facing_up: New papers (CVPR/ICCV/ECCV/JSTARS/TGRS/etc.)
• :computer: Open-source implementations
• :bar_chart: Dataset annotations
• :bookmark_tabs: Technical summaries

:book: Table of Contents

1. Technical Background
2. Latest Papers
3. Datasets
4. Codebases
5. Evaluation

:rocket: Technical Background

Referring Remote Sensing Image Segmentation (RRSIS) combines natural language descriptions with pixel-level understanding of aerial or satellite imagery. Key challenges include:

• Multiscale Objects: Buildings (1-100m) vs Vehicles (2-5m)
• Sensor Variations: 0.3m (WorldView) to 10m (Sentinel-2)
• Linguistic Ambiguity: "The red-roofed building northeast of the river"
• Occlusion Complexity: Partial visibility due to clouds/shadows (avg. 15-30% coverage in satellite imagery)
• Temporal Dynamics: Seasonal changes (e.g. snow cover) affecting object appearance
• Geometric Distortion: Parallax errors up to 5-10 pixels in oblique aerial views

:newspaper: Latest Papers

Segmentation Models

:floppy_disk: Datasets

:computer: Codebases

:chart_with_upwards_trend: Evaluation

Core Metrics

Basic IoU (Intersection over Union)

$$\text{IoU} = \frac{|P \cap G|}{|P \cup G|}$$

• Where ( P ) = Predicted segmentation, ( G ) = Ground truth
• Fundamental measure for pixel-wise segmentation accuracy

Generalized IoU (gIoU)

$$\text{gIoU} = \frac{1}{N} \sum_{i=1}^{N} \text{IoU}_i$$

• Calculated by averaging IoU scores across all test samples
• Reflects average performance on individual instances
• Primary metric (more robust to outliers)

Cumulative IoU (cIoU)

$$\text{cIoU} = \frac{\sum_{i=1}^{N} |P_i \cap G_i|}{\sum_{i=1}^{N} |P_i \cup G_i|}$$

• Computes ratio of cumulative intersections to unions
• Sensitive to large target areas (higher variance)

Precision@X (Pr@X)

Implementation

Standard metric implementation reference:
iou_metrics.py based on TorchMetrics.

---

:construction: Project under active development - Contribution Guidelines