README.md

---
language:
- en
license:
- cc-by-nc-4.0
task_categories:
- image-segmentation
splits:
  - name: train
  - name: validation
  - name: test
viewer: false
---

# OCTScenes: A Versatile Real-World Dataset of Tabletop Scenes for Object-Centric Learning 

This is the official dataset of OCTScenes (https://arxiv.org/abs//2306.09682). OCTScenes contains 5000 tabletop scenes with a total of 15 everyday objects. Each scene is captured in 60 frames covering a 360-degree perspective. 

In the OCTScenes-A dataset, the 0--3099 scenes without segmentation annotation are for training, while the 3100--3199 scenes with segmentation annotation can be used for testing. In the OCTScenes-B dataset, the 0--4899 scenes without segmentation annotation are for training, while the 4900--4999 scenes with segmentation annotation can be used for testing.

We have images of three different resolutions for each scene: 128x128, 256x256 and 640x480. The name of each image is in the form `[scene_id]_[frame_id].png`. They are in `./128x128`,`./256x256`, and `./640x480` respectively. The images are compressed using `tar`, and the name of compressed files starts with the resolutions, such as 'image_128x128_'. Please download all the compressed files, and use 'tar' instruction to decompress the files.

For example, for the images with the resolution of 128x128, please download all the files starts with 'image_128x128_', and then merge files into 'image_128x128.tar.gz':

```
cat image_128x128_* > image_128x128.tar.gz
```

And then decompress the file:

```
tar xvzf image_128x128.tar.gz
``` 

The segmentation results are in `./128x128/segments_128.tar.gz`. Currently, we have the segmentations of scenes 3100-3199 and 4900-4999 with resolutions 128x128. The shape of all of the images is 1x128x128, and the name of each segmentation is in the form `[scene_id]_[frame_id].png`. The int number in each pixel represents the index of the object (ranges from 1 to 10, and 0 represents the background).

The repository currently contains 128x128 size and 256x256 size datasets, and 640x480 size dataset will be available soon.