File size: 6,018 Bytes
25a49bc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
This is not an officially supported Google product.

# NAVI: Category-Agnostic Image Collections with High-Quality 3D Shape and Pose Annotations
### [Project Page](https://navidataset.github.io/) | [All objects](https://docs.google.com/presentation/d/1LCWUBQHs3oGi1bwCQjgLm8-etV9Y8ldPSpFN0BLoi6o/)

This repo contains a tutorial about how to download and use the NAVI dataset.


### Dataset versions 

#### v1.5 (latest)
Newly added annotated video scenes, for the same objects.
Download [here](https://storage.googleapis.com/gresearch/navi-dataset/navi_v1.5.tar.gz) (30GB).

#### v1.0
First release of in-the-wild and multiview collections.
Used in the experiments of the NeurIPS 2023 paper.
Download [here](https://storage.googleapis.com/gresearch/navi-dataset/navi_v1.0.tar.gz) (16GB).

## Overview of dataset contents.

The NAVI dataset consists of precise 3D-object-to-image alignments.
It contains:
- 36 object scans.
- 28921 precise object-to-image alignments in total.
- 324 (267 unique) multi-view scenes (8217 object-to-image alignments).
- 35 in-the-wild collections with different backgrounds (2298 object-to-image alignments).
- 176 (140 unique) video scenes (18406 object-to-image alignments).

There are three different types of images sets for each object:
- Multiview image collections (`multiview-XX-camera_model/`) capture the same object in the same environment and pose by moving the cameras.
- In-the-wild image collections (`wild_set/`) capture the same object under different illumination, background, pose.
- Video scenes (`video*-XX-camera_model/`) are similar to multiview collections, but are captured by a video, often with blurrier frames.

The folder organization is as follows:
```
object_id_0/
    3d_scan/
        object_id_0.obj
        object_id_0.mtl  # For textured objects.
        object_id_0.jpg  # For textured objects.
        object_id_0.glb
    multiview-00-camera_model/
        annotations.json
        images/
            000.jpg
            ...
        depth/  # Pre-computed depth.
            000.png
            ...
        masks/  # Pre-computed masks.
            000.png
            ...
    multiview-01-camera_model/
        ...
    ...
    video-00-camera_model/
        ...
        video.mp4
    ...
    wild_set/
        annotations.json
        images/
            000.jpg
            ...
        depth/
            ...
        masks/
            ...
object_id_1/
    ...
...
```

Each of the `annotations.json` contains the following information.
```
[
  {                                    # First object.
    "object_id": "3d_dollhouse_sink",  # The object id.
    "camera": {
      "q": [qw, qx, qy, qz],           # camera extrinsics rotation (quaternion).
      "t": [tx, ty, tz],               # camera extrinsics translation.
      "focal_length": 3024.0,          # Focal length in pixels.
      "camera_model": "pixel_5",       # Camera model name.
    },
    "filename": "000.jpg",             # The image file name under `images/`
    "filename_original": "PXL_20230304_014157778",  # The original image file name.
    "image_size": [3024, 4032],        # (height, width) of the image.
    "scene_name": "wild_set",          # The scene name that the image belongs to.
    "split": "train",                  # 'train' or 'val' split.
    "occluded": false,                 # Whether any part of the object is occluded.
    'video_id': 'MVI_2649'             # The video id, if applicable (for video scenes).
  },
  {...},                               # Second object.
  ...
]
```


## Download the dataset.

```bash
# Download (v1.5) 
wget https://storage.googleapis.com/gresearch/navi-dataset/navi_v1.5.tar.gz

## Links for previous versions.
# v1.0 
# wget https://storage.googleapis.com/gresearch/navi-dataset/navi_v1.0.tar.gz

# Extract
tar -xzf navi_v1.tar.gz
```

## Clone the code and use the dataset.

```bash
git clone https://github.com/google/navi
cd navi
```

### Use the dataset
Please refer to the included Notebook file `NAVI Dataset Tutorial.ipynb`.
Replace `/path/to/dataset/` with the correct directory.

Install the required packages:
```bash
conda create --name navi python=3
conda activate navi
pip install -r requirements.txt
```

Start Jupyter:
```bash
jupyter notebook
```


## Training and validation splits

We have released the common train/val splits in the json files.
In these splits, approximately 80% of the data under each folder was selected
for the train set, and 20% for the validation set.

Some tasks that are described in the paper have different setups, and require
different splits. We provide the splits that we used for these tasks in the
NeurIPS 2023 paper. 

#### For 3D from a single image.
For single-image 3D, there is the possibility to train and test a method on
different sets of object shapes. In that case, please use the object splits
provided under
`/path/to/navi/custom_splits/single_image_3d/objects-{train, val}.txt`.

In any other case, please use the default 80-20 splits of the json files.

#### For Pairwise pixel correspondence (sparse and dense)
Pairwise correspondences are evaluated on pairs of images.
We provide those under
`custom_splits/pairwise_pixel_correspondences/pairs-{multiview, wild_set}.txt`.

Each row is formatted as:

```
image_path_1 image_path_2 angular_distance_of_cameras
```
The angular distance is given in degrees, from 0 to 180.


## Citation

If you find this dataset useful, please consider citing our work:
```
@inproceedings{jampani2023navi,
  title={{NAVI}: Category-Agnostic Image Collections with High-Quality 3D Shape and Pose Annotations},
  author={Jampani, Varun and Maninis, Kevis-Kokitsi and Engelhardt, Andreas and Truong, Karen and Karpur, Arjun and Sargent, Kyle and Popov, Stefan and Araujo, Andre and Martin-Brualla, Ricardo and Patel, Kaushal and Vlasic, Daniel and Ferrari, Vittorio and Makadia, Ameesh and Liu, Ce and Li, Yuanzhen and Zhou, Howard},
  booktitle={NeurIPS},
  url={https://navidataset.github.io/},
  year={2023}
}
```