aadityabuilds/tree-distribution-shift

Tree Distribution Shift — Satellite Tree Detection (COCO + HF Datasets)

This is a dataset containing ~30K COCO tree crown annotated satellite image tiles of 400x400 px dimensions. These annotations come from all states in India and California in the United States. This dataset is organized as configs (distribution shift benchmarks). Each config provides three splits:

• train — 90% of the in-distribution pool
• id_test — 10% held-out from the in-distribution pool
• ood_test — out-of-distribution pool (base configs use the full OOD pool)

What is a config?

A config fully defines a benchmark setting (e.g., country shift, biome shift, region shift). When you select a config, you automatically get:

• train split
• id_test split
• ood_test split

No filtering or custom split logic is required.

Shared test sets

Configs come in pairs. The in-distribution split is always 90/10 train/id_test, and base configs use the full opposing pool as ood_test.

Config	train	id_test	ood_test
intl_train_IN__ood_US	90% of India	10% of India	all US
intl_train_US__ood_IN	90% of US	10% of US	all India

Few-shot variants (__fs1, __fs10, __fs100, __fsall) move K images from ood_test into train.

Available base configs (8)

#	Config name	Shift type	Train on	OOD from
1	intl_train_IN__ood_US	Country	India	US
2	intl_train_US__ood_IN	Country	US	India
3	biome_Rajasthan_train_WET__ood_DRY	Biome	Rajasthan WET	Rajasthan DRY
4	biome_Rajasthan_train_DRY__ood_WET	Biome	Rajasthan DRY	Rajasthan WET
5	elev_Karnataka_train_HIGH__ood_LOW	Elevation	Karnataka HIGH	Karnataka LOW
6	elev_Karnataka_train_LOW__ood_HIGH	Elevation	Karnataka LOW	Karnataka HIGH
7	region_train_North__ood_South	Region	North India	South India
8	region_train_South__ood_North	Region	South India	North India

For each base config, there are 4 few-shot variants:

• __fs1
• __fs10
• __fs100
• __fsall

Total configs = 8 base + 32 few-shot = 40.

---

Export to COCO (recommended for training)

Use this when you want a standard COCO folder structure that works with most CV frameworks (Detectron2, MMDetection, torchvision detection, etc.).

Step 1 — Clone the dataset repository (to get the export tool)

git clone https://huggingface.co/datasets/aadityabuilds/tree-distribution-shift
cd tree-distribution-shift

Step 2 — Install dependencies

pip install datasets huggingface_hub orjson

Step 3 — Pick a config

Go to the dataset page and choose a config from the Configs section:

https://huggingface.co/datasets/aadityabuilds/tree-distribution-shift

Example config:

intl_train_IN__ood_US

Step 4 — Export to COCO (one command)

python tools/export_coco.py \
  --repo aadityabuilds/tree-distribution-shift \
  --config intl_train_IN__ood_US \
  --out ./coco_out

Output structure

coco_out/
└── intl_train_IN__ood_US/
    ├── train/
    │   ├── images/
    │   │   └── *.tiff
    │   └── annotations/
    │       └── instances_train.json
    ├── id_test/
    │   ├── images/
    │   └── annotations/
    │       └── instances_id_test.json
    └── ood_test/
        ├── images/
        └── annotations/
            └── instances_ood_test.json

---

Load with Hugging Face Datasets (recommended for analysis / prototyping)

Use this when you want programmatic access (e.g., notebooks, statistics, custom pipelines) without writing files to disk.

Install

pip install datasets

Load a config (gets train / id_test / ood_test)

from datasets import load_dataset

ds = load_dataset(
    "aadityabuilds/tree-distribution-shift",
    "intl_train_IN__ood_US"
)

train    = ds["train"]
id_test  = ds["id_test"]
ood_test = ds["ood_test"]

print(ds)
print(len(train), len(id_test), len(ood_test))
print(train[0].keys())

What each example contains

Each row includes:

• image_bytes (raw image bytes)
• coco_annotations (COCO annotations for that image)
• coco_categories (COCO categories)
• metadata fields (country, state, zone, region, biome, etc.)