---
license: mit
task_categories:
  - other
tags:
  - 3d
  - animation
  - mesh
  - objaverse
  - 4d
size_categories:
  - 10K<n<100K
---

# DynamicObjaverseProcessed

67,084 animated (4D) 3D meshes curated from [Objaverse-XL](https://objaverse.allenai.org/), quality-ranked by ICP-based non-rigid deformation intensity.

## Overview

Starting from 71,104 curated UIDs, the [processing pipeline](https://github.com/ou524u/DynamicObjaverseProcessing) downloads, unifies formats to `.glb`, and classifies meshes by animation quality:

```
71,104 curated UIDs → 67,746 downloads → 67,639 unified .glb → 67,084 true 4D meshes
```

All meshes are in binary glTF (`.glb`) format with embedded animations.

## Quintile Ranking

Meshes are ranked by **ICP residual** — a measure of non-rigid deformation after factoring out rigid-body motion (translation/rotation). Higher ICP = more interesting mesh deformation (e.g., character animation, cloth simulation).

| Quintile | Files | ICP Range | Size | Description |
|----------|-------|-----------|------|-------------|
| Q1 (top) | 13,417 | 0.039 – 0.996 | 32 GB | Strongest deformation |
| Q2 (high) | 13,417 | 0.027 – 0.039 | 44 GB | |
| Q3 (mid) | 13,417 | 0.017 – 0.027 | 51 GB | |
| Q4 (low) | 13,417 | 0.003 – 0.017 | 38 GB | |
| Q5 (bottom) | 13,416 | 0.000 – 0.003 | 25 GB | Subtle / rigid-body motion |

## Files

- `classification.jsonl` — Full manifest with per-mesh metrics (ICP residual, amplitude, quintile, rank)
- `quintile_1_top.tar` — Top 20% by deformation intensity (13,417 `.glb` files)
- `quintile_2_high.tar` — High 20%
- `quintile_3_mid.tar` — Mid 20%
- `quintile_4_low.tar` — Low 20%
- `quintile_5_bottom.tar` — Bottom 20%

## Usage

```python
# Download and extract a specific quintile
from huggingface_hub import hf_hub_download
import tarfile

path = hf_hub_download("littlekoyo/DynamicObjaverseProcessed", "quintile_1_top.tar", repo_type="dataset")
with tarfile.open(path) as tar:
    tar.extractall("./meshes")

# Load the classification manifest
import json
path = hf_hub_download("littlekoyo/DynamicObjaverseProcessed", "classification.jsonl", repo_type="dataset")
records = [json.loads(line) for line in open(path)]
```

### classification.jsonl fields

| Field | Description |
|-------|-------------|
| `uid` | UUID identifier (also the filename without `.glb`) |
| `classification` | `true_4d` / `static` / `error` |
| `icp_residual_normalized` | Non-rigid deformation intensity (normalized by bbox diagonal) |
| `max_displacement_normalized` | Raw vertex displacement (including rigid motion) |
| `overall_amplitude` | Keyframe amplitude metric (supplementary) |
| `quintile` | 1–5 (1 = most deformation, 5 = least; 0 for non-4D) |
| `icp_rank` | Rank within true_4d (1 = highest ICP) |
| `source_path` | Original file path in the processing pipeline |

## Processing Pipeline

See [DynamicObjaverseProcessing](https://github.com/ou524u/DynamicObjaverseProcessing) for the full pipeline code.

## License

The pipeline code is MIT licensed. Mesh data is sourced from Objaverse-XL and subject to its original licenses.