DannHiroaki/COCO-Spatial-Join-1.23B

---
license: cc-by-4.0
tags:
- coco
- sptial-join

configs:
- config_name: rects
default: true
data_files:
- split: train
path: "data/rects/train2017/*.parquet"
- split: validation
path: "data/rects/val2017/*.parquet"

- config_name: images
data_files:
- split: train
path: "data/images.parquet"
---

# Inroduction

**COCO-Spatial-Join-1B** is a large-scale, deterministic **spatial join benchmark** constructed from the **MS COCO 2017** detection annotations and **RPN proposals** produced by **Detectron2 Faster R-CNN (ResNet-50-FPN)**. The benchmark is designed to stress-test spatial join systems under **high-overlap** workloads while providing an unambiguous geometric semantics.

All objects (ground-truth and proposals) are represented as **axis-aligned half-open 3D boxes** in a shared coordinate system. A **global spatial join** can be evaluated over the complete corpus, while the z-dimension construction yields a clean per-image decomposition when desired.

Dataset construction details and the reference builder are available at:
https://github.com/DANNHIROAKI/COCO-Spatial-Join-1B-Builder

# Example

Installation

```shell
pip install -U huggingface_hub
```

Download the Entire Dataset

```shell
hf download DannHiroaki/COCO-Spatial-Join-1.23B \
--repo-type dataset \
--local-dir ./COCO-Spatial-Join-1.23B
```

Download specific shards from `train2017`

```shell
hf download DannHiroaki/COCO-Spatial-Join-1.23B \
--repo-type dataset \
data/rects/train2017/shard-000000.parquet \
data/rects/train2017/shard-001024.parquet \
--local-dir ./COCO-Spatial-Join-rects-sample

```

Dry Run (Check size before downloading)

```shell
hf download DannHiroaki/COCO-Spatial-Join-1.23B --repo-type dataset --include "data/rects/val2017/*.parquet" --dry-run
```