Hugging Face's logo

Datasets:
aadarsh99
/
kyvo-datasets-and-codebooks

Modalities:
Image
License:
Dataset card Data Studio Files Community
2
aadarsh99 commited on
Commit
24761db
Β·
1 Parent(s): ebdd26e

added README

Browse files
Files changed (1) hide show
  1. README.md +87 -0
README.md ADDED
@@ -0,0 +1,87 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # 3D MLLM Dataset and Codebooks Details
2
+
3
+ This document provides details about the dataset and codebooks provided in the `3d-mllm-datasets-and-codebooks` repository. We will provide the details about each of the folders in the repository and the contents of each folder.
4
+
5
+ ## Data Generation Pipeline
6
+
7
+ The pipeline that we follow to generate the pre-tokenized data is as follows:
8
+
9
+ * **3D Scenes**: 3D Scene JSON --> Serialized 3D Scene --> Tokenized 3D Scene
10
+ * **Images**: Image --> VQGAN Codebook Indices --> Tokenized Image
11
+ * **Text**: Text --> Tokenized Text
12
+
13
+
14
+
15
+
16
+ ## Pre-tokenized Data
17
+
18
+ The `pretokenized-data` folder contains all the pre-tokenized data for the datasets used in the 3D MLLM project. The pre-tokenized data is stored in the following structure:
19
+
20
+ ```python
21
+ pretokenized-data/
22
+ |-- clevr/
23
+ | |-- 3d-scenes/ # contains all pre-tokenized 3D scenes for CLEVR for all tasks
24
+ | |-- images/ # contains all pre-tokenized images for CLEVR for all tasks
25
+ | |-- text/ # contains all pre-tokenized text for CLEVR for all tasks
26
+ |-- objaworld/
27
+ | |-- 3d-scenes/ # contains all pre-tokenized 3D scenes for ObjaWorld for all tasks
28
+ | |-- images/ # contains all pre-tokenized images for ObjaWorld for all tasks
29
+ |-- objectron/
30
+ | |-- 3d-scenes/ # contains all pre-tokenized 3D scenes for Objectron for all tasks
31
+ | |-- images/ # contains all pre-tokenized images for Objectron for all tasks
32
+ ```
33
+
34
+
35
+ For a given task, an input can be any combination of 3d-scenes, images, and text. The output can be any combination of images, text, and 3d-scenes. In the following table we outline the tasks for each dataset and the corresponding input and output data that are needed for each task.
36
+
37
+ | **Task** | **Input Image** | **Input 3D Scene** | **Input Text** | **Output Image** | **Output 3D Scene** | **Output Text** |
38
+ |:----------------------:|:------------------:|:----------------------:|:-----------------:|:------------------:|:-----------------------:|:-----------------:|
39
+ | **CLEVR** | | | | | | |
40
+ | Rendering | 𐄂 | βœ“ | 𐄂 | βœ“ | 𐄂 | 𐄂 |
41
+ | Recognition | βœ“ | 𐄂 | 𐄂 | 𐄂 | βœ“ | 𐄂 |
42
+ | Instruction-Following | βœ“ | βœ“ | βœ“ | βœ“ | βœ“ | 𐄂 |
43
+ | Question-Answering | βœ“ | βœ“ | βœ“ | 𐄂 | 𐄂 | βœ“ |
44
+ | | | | | | | |
45
+ | **ObjaWorld** | | | | | | |
46
+ | Rendering | 𐄂 | βœ“ | 𐄂 | βœ“ | 𐄂 | 𐄂 |
47
+ | Recognition | βœ“ | 𐄂 | 𐄂 | 𐄂 | βœ“ | 𐄂 |
48
+ | | | | | | | |
49
+ | **Objectron** | | | | | | |
50
+ | Recognition | βœ“ | 𐄂 | 𐄂 | 𐄂 | βœ“ | 𐄂 |
51
+
52
+ For the exact files that correspond to the input and output data for each task, please refer to the corresponding configuration files in the `configs/llama3_2/train` folder.
53
+
54
+
55
+ ## VQGAN Models and Codebooks
56
+
57
+ The `vqgan-models-and-codebooks` folder contains all the VQGAN model checkpoints and codebooks for the datasets used in the 3D MLLM project. The VQGAN model checkpoints and codebooks are stored in the following structure:
58
+
59
+ ```python
60
+ vqgan-models-and-codebooks/
61
+ |-- clevr/
62
+ | |-- 2024-10-10T09-21-36_custom_vqgan_CLEVR-LARGE/ # contains the VQGAN model checkpoint for CLEVR
63
+ | |-- custom_vqgan_embedding_1024CLEVRLARGE_256dim.npy # contains the VQGAN codebook for CLEVR
64
+ |-- domain-agnostic/
65
+ | |-- vqgan_gumbel_f8/ # contains the VQGAN model checkpoint for Domain Agnostic VQGAN (provided by taming-transformers)
66
+ | |-- quantize_weight_8192.npy # contains the VQGAN codebook for Domain Agnostic VQGAN
67
+ |-- objaworld/
68
+ | |-- 2025-01-17T09-02-22_custom_vqgan_SYNTHETIC_LIVINGROOM_PARK_LARGE_EP100/ # contains the VQGAN model checkpoint for ObjaWorld
69
+ | |-- custom_vqgan_embedding_256SYNTHETIC_LIVINGROOM_PARK_LARGE_EP100_256dim.npy # contains the VQGAN codebook for ObjaWorld
70
+ |-- objectron/
71
+ | |-- 2024-11-03T05-41-42_custom_vqgan_OMNI3D_OBJECTRON_ep200/ # contains the VQGAN model checkpoint for Objectron
72
+ | |-- custom_vqgan_embedding_256Omni3D-OBJECTRON_256dim.npy # contains the VQGAN codebook for Objectron
73
+ ```
74
+
75
+ ## Images and Scenes for Evaluation
76
+
77
+ The `images-and-scenes-for-evaluation` folder contains all the groundtruth images and scenes for the datasets used in the 3D MLLM project. The images and scenes are used to compute the evaluation metrics for the different tasks. The images and scenes are stored in the following structure:
78
+
79
+ ```python
80
+ images-and-scenes-for-evaluation/
81
+ |-- clevr/ # contains all images and scenes for evaluation for CLEVR
82
+ |-- objaworld/ # contains all images and scenes for evaluation for ObjaWorld
83
+ |-- objectron/ # contains all scenes for evaluation for Objectron
84
+ ```
85
+
86
+
87
+