File size: 4,585 Bytes
5c43d50
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a883bd3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
97787d4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5c43d50
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
---
license: cc-by-4.0
task_categories:
  - object-detection
  - keypoint-detection
  - image-segmentation
tags:
  - synthetic
  - SEM
  - scanning-electron-microscopy
  - materials-science
  - ZIF-8
  - metal-organic-framework
  - 3d-keypoint
  - instance-segmentation
size_categories:
  - 10K<n<100K
pretty_name: SynZIF-8
---

# SynZIF-8: A Synthetic SEM Dataset and Benchmark for 3D Sub-Micron Crystal Perception

## Dataset Summary

SynZIF-8 is a synthetic Scanning Electron Microscope (SEM) dataset of 
rhombic dodecahedron crystals modeled after ZIF-8 metal-organic frameworks. 
The dataset contains **20,000 images** with **455K+ instance annotations** 
and zero human annotation cost, generated by chaining ControlNet and 
LoRA-stylized diffusion models with deterministic 3D rendering.

## Code Repository

The evaluation code, training scripts, and baseline implementations referenced
in the SynZIF-8 paper (NeurIPS 2026) are released at:

**🔗 https://github.com/synzif8/synzif8**

This includes:
- 3D rendering pipeline (rhombic dodecahedron geometry, Gaussian clustering,
  orthographic projection, per-instance annotation extraction)
- SEM stylization (Stable Diffusion 1.5 + LoRA + ControlNet + mask-guided
  contrast calibration)
- Mask R-CNN front-end detector (seg_exp07)
- Nine baseline implementations across three families:
  - 3D keypoint regression: FFB6D, REDE, Uni6D
  - Geometry-aware 6D pose: SC6D, GDR-Net, HccePose
  - Foundation-style 6D pose: MegaPose, GigaPose, FoundationPose
- Evaluation suite: edge-length RRMSE, MPJPE with per-instance z-offset
  correction
- FastAPI annotation tool + chemistry expert annotation notebooks

## Trained Baseline Checkpoints

Pre-trained weights for all nine baseline models reported in the paper (Sec. 4)
are provided under `checkpoints/`:

| File | Model | Category | Size |
|------|-------|----------|------|
| `checkpoints/ffb6d.pt` | FFB6D | (B) 3D Keypoint Regression | 141 MB |
| `checkpoints/rede.pt` | REDE | (B) 3D Keypoint Regression | 146 MB |
| `checkpoints/uni6d.pt` | Uni6D | (B) 3D Keypoint Regression | 606 MB |
| `checkpoints/sc6d.pt` | SC6D | (A) Geometry-Aware 6D Pose | 146 MB |
| `checkpoints/gdrnet.pt` | GDR-Net | (A) Geometry-Aware 6D Pose | 155 MB |
| `checkpoints/hccepose.pt` | HccePose | (A) Geometry-Aware 6D Pose | 112 MB |
| `checkpoints/megapose.pt` | MegaPose | (C) Foundation-Style 6D Pose | 120 MB |
| `checkpoints/gigapose.pt` | GigaPose | (C) Foundation-Style 6D Pose | 583 MB |
| `checkpoints/foundationpose.pt` | FoundationPose | (C) Foundation-Style 6D Pose | 17 MB |

Each file is a PyTorch `state_dict` (pure model weights, no metadata).
Load with `torch.load("checkpoints/<model>.pt", map_location="cpu")`.

## Dataset Structure

The dataset is provided as a single tar archive (`synzif8_full.tar`) 
containing the `dataset_v6/` directory.

### Per-image files

For each rendered scene `render_{ID}`, the following files are provided:

| File | Description |
|------|-------------|
| `render_{ID}.png` | Original rendered image (rhombic dodecahedron crystals) |
| `render_{ID}_styled.png` | SEM-stylized image (via ControlNet + LoRA) |
| `render_{ID}_edge.png` | Visible edge map |
| `render_{ID}_labeled.png` | Labeled visualization with instance IDs |
| `render_{ID}_mask.png` | Combined instance mask |
| `render_{ID}_masks/` | Per-instance amodal and visible masks |
| `render_{ID}_metadata.json` | 14 vertex coordinates, edge lengths, visibility, per-instance annotations |

## Splits

| Split | # Images |
|-------|---------:|
| Train | 16,000 |
| Validation | 2,000 |
| Test | 2,000 |

Split is fixed (seed=42) and shared across all benchmark experiments.

## Tasks Supported

- **Instance Segmentation** (amodal + visible masks)
- **3D Keypoint Detection** (14 vertices per crystal)
- **6D Pose Estimation**
- **Edge Length Measurement** (for downstream CO₂ diffusivity computation)

## Intended Use

SynZIF-8 is designed as both training data and an evaluation testbed for:
1. Models tackling SEM image analysis of crystalline particles
2. Benchmarking automated morphological measurement (edge length) against 
   human expert annotation
3. Evaluating 3D keypoint detection and 6D pose estimation in 
   sub-micron-scale microscopy domains

## License

CC-BY-4.0

## Limitations

- Synthetic-to-real domain gap may persist despite SEM stylization
- Models exclusively cover ZIF-8-like rhombic dodecahedron geometry
- A 2-pixel resolution discrepancy exists between rendered (686px) and 
  SEM-styled (688px) images

## Citation

To be added upon publication.