PeterAM4 commited on
Commit
0b8cbca
·
verified ·
1 Parent(s): 0ba25b9

Update Readme

Browse files
Files changed (1) hide show
  1. README.md +154 -0
README.md CHANGED
@@ -54,3 +54,157 @@ configs:
54
  - split: test
55
  path: data/test-*
56
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
54
  - split: test
55
  path: data/test-*
56
  ---
57
+
58
+ # BlockGen-3D Dataset
59
+
60
+ ## Overview
61
+
62
+ BlockGen-3D is a large-scale dataset of voxelized 3D models with accompanying text descriptions, specifically designed for text-to-3D generation tasks. By processing and voxelizing models from the [Objaverse dataset](https://huggingface.co/datasets/allenai/objaverse), we have created a standardized representation that is particularly suitable for training 3D generative models.
63
+
64
+ Our dataset provides two types of representations: shape-only models represented as binary occupancy grids, and colored models with full RGBA information. Each model is represented in a 32×32×32 voxel grid, striking a balance between detail preservation and computational efficiency. This uniform representation makes the dataset especially suitable for training diffusion models and other deep learning architectures for 3D generation.
65
+
66
+ The dataset contains 542,292 total samples, with:
67
+ - 515,177 training samples
68
+ - 27,115 test samples
69
+
70
+ ## Data Format and Structure
71
+
72
+ Every sample in our dataset consists of a voxel grid with accompanying metadata. The voxel grid uses a consistent 32³ resolution, with values structured as follows:
73
+
74
+ For shape-only data:
75
+ - `voxels_occupancy`: A binary grid of shape [1, 32, 32, 32] where each voxel is either empty (0) or occupied (1)
76
+
77
+ For colored data:
78
+ - `voxels_colors`: RGB color information with shape [3, 32, 32, 32], normalized to [0, 1]
79
+ - `voxels_occupancy`: The occupancy mask as described above
80
+
81
+ Each sample also includes rich metadata:
82
+ - Text descriptions derived from the original Objaverse dataset
83
+ - Categorization and tagging information
84
+ - Augmentation status and original file information
85
+ - Number of occupied voxels for quick filtering or analysis
86
+
87
+ ## Using the Dataset
88
+
89
+ ### Basic Loading and Inspection
90
+
91
+ ```python
92
+ from datasets import load_dataset
93
+
94
+ # Load the dataset
95
+ dataset = load_dataset("PeterAM4/blockgen-3d")
96
+
97
+ # Access splits
98
+ train_dataset = dataset["train"]
99
+ test_dataset = dataset["test"]
100
+
101
+ # Examine a sample
102
+ sample = train_dataset[0]
103
+ print(f"Sample description: {sample['name']}")
104
+ print(f"Number of occupied voxels: {sample['metadata']['num_occupied']}")
105
+ ```
106
+
107
+ ### Working with Voxel Data
108
+
109
+ When working with the voxel data, you'll often want to handle both shape-only and colored samples uniformly. Here's a utility function that helps standardize the processing:
110
+
111
+ ```python
112
+ import torch
113
+
114
+ def process_voxel_data(sample, default_color=[0.5, 0.5, 0.5]):
115
+ """Process voxel data with default colors for shape-only data.
116
+
117
+ Args:
118
+ sample: Dataset sample
119
+ default_color: Default RGB values for shape-only models (default: gray)
120
+
121
+ Returns:
122
+ torch.Tensor: RGBA data with shape [4, 32, 32, 32]
123
+ """
124
+ occupancy = torch.from_numpy(sample['voxels_occupancy'])
125
+
126
+ if sample['voxels_colors'] is not None:
127
+ # Use provided colors for RGBA samples
128
+ colors = torch.from_numpy(sample['voxels_colors'])
129
+ else:
130
+ # Apply default color to shape-only samples
131
+ default_color = torch.tensor(default_color)[:, None, None, None]
132
+ colors = default_color.repeat(1, 32, 32, 32) * occupancy
133
+
134
+ # Combine into RGBA format
135
+ rgba = torch.cat([colors, occupancy], dim=0)
136
+ return rgba
137
+ ```
138
+
139
+ ### Batch Processing with DataLoader
140
+
141
+ For training deep learning models, you'll likely want to use PyTorch's DataLoader. Here's how to set it up with a simple prompt strategy:
142
+
143
+ For basic text-to-3D generation, you can use the model names as prompts:
144
+ ```python
145
+ def collate_fn(batch):
146
+ """Simple collate function using basic prompts."""
147
+ return {
148
+ 'voxels': torch.stack([process_voxel_data(x) for x in batch]),
149
+ 'prompt': [x['name'] for x in batch] # Using name field as prompt, could also use more complex prompts from categories and tags dict keys
150
+ }
151
+
152
+ # Create DataLoader
153
+ dataloader = DataLoader(
154
+ train_dataset,
155
+ batch_size=32,
156
+ shuffle=True,
157
+ collate_fn=collate_fn
158
+ )
159
+ ```
160
+
161
+ ## Visualization Examples
162
+
163
+
164
+
165
+ ## Dataset Creation Process
166
+
167
+ Our dataset was created through the following steps:
168
+
169
+ 1. Starting with the Objaverse dataset
170
+ 2. Voxelizing 3D models at 32³ resolution
171
+ 3. Preserving color information where available
172
+ 4. Generating data augmentations through rotations
173
+ 5. Creating train/test splits (95%/5% split ratio)
174
+
175
+ ## Training tips
176
+
177
+ - For shape-only models, use only the occupancy channel
178
+ - For colored models:
179
+
180
+ - Apply colors only to occupied voxels
181
+ - Use default colors for shape-only samples
182
+
183
+ ## Citation
184
+
185
+ If you use this dataset in your research, please cite:
186
+
187
+ @misc{blockgen2024,
188
+ title={BlockGen-3D: A Large-Scale Dataset for Text-to-3D Voxel Generation},
189
+ author={Peter A. Massih},
190
+ year={2024},
191
+ publisher={Hugging Face}
192
+ }
193
+
194
+ Please also cite the original Objaverse dataset:
195
+
196
+ @article{objaverse2023,
197
+ title={Objaverse: A Universe of Annotated 3D Objects},
198
+ author={Matt Deitke and Dustin Schwenk and Jordi Salvador and Luca Weihs and Oscar Michel and Eli VanderBilt and Ludwig Schmidt and Kiana Ehsani and Aniruddha Kembhavi and Ali Farhadi},
199
+ journal={arXiv preprint arXiv:2304.02643},
200
+ year={2023}
201
+ }
202
+
203
+ ## Limitations
204
+
205
+ - Fixed 32³ resolution limits fine detail representation
206
+ - Not all models have color information
207
+ - Augmentations are limited to rotations
208
+ - Voxelization may lose some geometric details
209
+
210
+