SelvaMask: Segmenting Trees in Tropical Forests and Beyond

BCI50ha	Detectree2	SelvaMask (ours)	Predictions (ours)

Qualitative comparison between SelvaMask and other tree crown segmentation datasets.

Dataset Statistics

The table below details the annotated area (in hectares) and individual crown counts across geographic sites and splits.

Site	Train Area (ha)	Train Crowns	Val Area (ha)	Val Crowns	Test Area (ha)	Test Crowns
Panama	5.15	1499	1.53	570	2.93	783
Brazil	3.31	1590	0.92	568	1.16	716
Ecuador	7.20	1836	1.89	629	2.91	896
Total	15.7	4925	4.3	1767	7.0	2395

Tiling Strategy & Split Counts

You may notice that the absolute number of images (tiles) in the train split is lower than expected compared to the validation and test sets, despite the training set covering the largest area (15.7 ha). This is due to our tiling strategy:

• Training: Tiles are extracted at 3555 × 3555 pixels with 50% overlap to maximize spatial context and support data augmentation.
• Validation & Test: Tiles are generated at 1777 × 1777 pixels with 75% overlap for standardized evaluation.

Dataset Structure

The dataset consists of TIFF images (accessible via PIL) and COCO-formatted annotations.

Downloading the dataset

Run the following script to fetch and print the first row in the training split:

from datasets import load_dataset
import matplotlib.pyplot as plt

# Load in streaming mode to verify structure without full download
ds_stream = load_dataset("selvamask/SelvaMask", split="train", streaming=True)
sample = next(iter(ds_stream))

# Print metadata keys and structure
print("Available keys:", sample.keys())
print("Metadata summary:", {k: v for k, v in sample.items() if k != 'image'})

# Visualise the first tile
plt.imshow(sample["image"])
plt.axis("off")
plt.title(f"Sample: {sample['tile_name']}", fontsize=10)
plt.show()