README.md

---
license: mit
---

# Streaming3D Dataset

This dataset contains assets used by the Streaming3D benchmark. The current
release documents the `GSO30` subset; other subsets may be added later.

## GSO30

`GSO30` is a 30-object subset derived from Google Scanned Objects. Each object
directory contains training renders, evaluation assets, and the original object
mesh/material files.

### Object List

```text
alarm backpack bell blocks chicken cream elephant grandfather grandmother hat
leather lion lunch_bag mario oil school_bus1 school_bus2 shoe shoe1 shoe2
shoe3 soap sofa sorter sorting_board stucking_cups teapot toaster train turtle
```

### Directory Structure

```text
GSO30/
  <object_id>/
    meshes/
      model.glb
      model.obj
      model.mtl
      texture.png
    render_spiral_100/
      images/
        000.png ... 099.png
      masks/
        000.png ... 099.png
      model/
        000.png ... 099.png
        000.npy ... 099.npy
      transforms.json
      model_norm.obj
      model_norm.mtl
    render_mvs_25/
      model_norm.glb
      model_norm.obj
      model_norm.mtl
      model/
        000.png ... 024.png
        000.npy ... 024.npy
```

Some object folders also include auxiliary metadata, thumbnails, or legacy
render folders. The benchmark protocol uses the paths above.

### Usage

For training or reconstruction input, use all 100 images from:

```text
GSO30/<object_id>/render_spiral_100/images/{000..099}.png
```

The corresponding masks are stored in:

```text
GSO30/<object_id>/render_spiral_100/masks/{000..099}.png
```

Camera metadata for the 100 spiral views is available in:

```text
GSO30/<object_id>/render_spiral_100/transforms.json
GSO30/<object_id>/render_spiral_100/model/{000..099}.npy
```

For evaluation, use the normalized GLB mesh and the 25 provided camera views
from `render_mvs_25`:

```text
GSO30/<object_id>/render_mvs_25/model_norm.glb
GSO30/<object_id>/render_mvs_25/model/{000..024}.npy
```

The matching reference renders for those views are:

```text
GSO30/<object_id>/render_mvs_25/model/{000..024}.png
```

In short, the default protocol is:

1. Train or reconstruct from all `render_spiral_100/images` frames.
2. Evaluate by rendering or comparing against `render_mvs_25/model_norm.glb`
   using the 25 camera poses in `render_mvs_25/model/*.npy`.