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

Dataset Summary

The OCTScenes dataset is a versatile real-world dataset of tabletop scenes for object-centric learning, containing 5000 tabletop scenes with a total of 15 objects. Each scene is captured in 60 frames covering a 360-degree perspective. It can satisfy the evaluation of object-centric learning methods based on single-image, video, and multi-view.

The 15 distinct types of objects are shown in Figure 1, and some examples of data are shown in Figure 2.

Figure 1: Objects of the dataset.

Figure 2: Examples of images, depth maps, and segmentation maps of the dataset.

Supported Tasks and Leaderboards

• object-centric learning: The dataset can be used to train a model for object-centric learning, which aims to learn compositional scene representations in an unsupervised manner. The segmentation performance of model is measured by Adjusted Mutual Information (AMI), Adjusted Rand Index (ARI), and mean Intersection over Union (mIoU). Two variants of AMI and ARI are used to evaluate the segmentation performance more thoroughly. AMI-A and ARI-A are computed using pixels in the entire image and measure how accurately different layers of visual concepts (including both objects and the background) are separated. AMI-O and ARI-O are computed only using pixels in the regions of objects and focus on how accurately different objects are separated. The reconstruction performance of model is measured by Minimize Squared Error (MSE) and Learned Perceptual Image Patch Similarity (LPIPS). Success on this task is typically measured by achieving high AMI, ARI, and mIOU and low MSE and LPIPS.

Languages

English.

Dataset Structure

We provide images of three different resolutions for each scene: 640x480, 256x256, and 128x128. The name of each image is in the form [scene_id]_[frame_id].png. They are available in ./640x480, ./256x256, and ./128x128, respectively.

The images are compressed using tar and the names of the compressed files start with the resolutions, e.g. image_128x128_. Please download all compressed files and use the tar command to decompress them.

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

cat image_128x128_* > image_128x128.tar.gz

And then decompress the file:

tar xvzf image_128x128.tar.gz

Data Instances

Each data instance contains an RGB image, its depth map, its camera intrinsic matrix, its camera pose, and its segmentation map, which is None in the training and validation set.

Data Fields

• scene_id: a string scene identifier for each example
• frame_id: a string frame identifier for each example
• resolution: a string for the image resolution of each example (e.g. 640x480, 256x256, 128x128)
• image: a PIL.Image.Image object containing the image
• depth: a PIL.Image.Image object containing the depth map
• segment: a PIL.Image.Image object containing the segmentation map, where the int number in each pixel represents the index of the object (ranges from 1 to 10, with 0 representing the background).
• intrinsic_matrix: a numpy.ndarray for the camera intrinsic matrix of each image
• camera_pose: a numpy.ndarray for the camera pose of each image

Data Splits

The data is split into two subsets to create datasets with different levels of difficulty levels of difficulty. Both the two subsets are randomly divided into training, validation, and testing sets. The validation and testing sets each consist of 100 scenes, while the remaining scenes form the training set. Only the data in the testing set contain segmentation annotations for evaluation.

OCTScenes-A contains 3200 scenes (scene_id from 0000 to 3199) and includes only the first 11 object types, with scenes consisting of 1 to 6 objects, making it comparatively smaller and less complex. Images with scene_id ranging from 0000 to 2999 are used for training, images with scene_id ranging from 3000 to 3099 are for validation, and images with scene_id ranging from 3100 to 3199 are for testing.

OCTScenes-A contains 5000 scenes (scene_id from 0000 to 4999) and includes all 15 object types, with scenes consisting of 1 to 10 objects, resulting in a larger and more complex dataset. Images with scene_id ranging from 0000 to 4799 are used for training, images with scene_id ranging from 4800 to 4899 are for validation, and images with scene_id ranging from 4900 to 4999 are for testing.

Dataset	OCTScenes-A			OCTScenes-B
Resolution	640x480	256x256	128x128	640x480	256x256	128x128
Split	train	validation	test	train	validation	test
Number of scenes	3000	100	100	4800	100	100
Number of object catergories	11			15
Number of objects in a scene	1~6			1~10
Number of views in a scene	60			60

Dataset Creation

Curation Rationale

OCTScenes was designed as a novel benchmark for unsupervised object-centric learning. It serves as a versatile real-world dataset that aims to fill the scarcity of specifically tailored real-world datasets in this field.

Source Data

Initial Data Collection and Normalization

A three-wheel omnidirectional wheel robot equipped with an Orbbec Astra 3D camera was employed for data collection. It took place in a school conference room, where a small wooden table was positioned on the floor and surrounded by baffles. Randomly selected objects, ranging from 1 to 10, were manually placed on the table without any stacking. The data was directly collected from these visual scenes.

Annotations

Annotation process

• Segmentation Annotation: We use EISeg, which is a high-performance interactive automatic annotation tool for image segmentation, to label the segmentation maps. We manually labeled 6 images of each scene and used the labeled images to train a supervision real-time semantic segmentation model named PP-LiteSeg using the framework PaddleSeg to annotation the rest of the data. The annotated images are split into 90% for training and 10% validation, achieving a mean Intersection over Union (mIoU) of 0.92 on the validation set.
• Intrinsic Matrix: We obtained the intrinsic matrix of the camera through camera calibration.
• Camera Pose: We obtained the camera pose of each image through 3D reconstruction using COLMAP, which is commonly used to create real-world NeRF datasets.

Who are the annotators?

Some annotations are manually labelled by the authors, while others are generated by the model.

Personal and Sensitive Information

N/A

Considerations for Using the Data

Social Impact of Dataset

N/A

Discussion of Biases

N/A

Other Known Limitations

The main limitation of the dataset is its simplicity, characterized by a single background type and uncomplicated object shapes, most of which are symmetrical and lack the variation in orientation that occurs when viewed from different perspectives. Therefore, the object representations learned by the model are relatively simple, and some simple modeling methods may produce better segmentation results than complex modeling methods.

To overcome the aforementioned issue and enhance the dataset further, we have devised a plan for the next version of OCTScenes. In our future work, we will introduce a wider range of diverse and complex backgrounds, including tables of different types, patterns, and materials, and a greater variety of objects into the OCTScenes, particularly objects with asymmetric shapes, complex textures, and mixed colors, which will increase the complexity and diversity of the dataset.

Additional Information

Dataset Curators

The dataset was created by Yinxuan Huang, Tonglin Chen, Zhimeng Shen, Jinghao Huang, Bin Li, and Xiangyang Xue as members of the Visual Intelligence Lab at Fudan University.

Licensing Information

The dataset is available under CC-BY-NC 4.0 license.

Citation Information

@article{huang2023octscenes,
  title={OCTScenes: A Versatile Real-World Dataset of Tabletop Scenes for Object-Centric Learning},
  author={Huang, Yinxuan and Chen, Tonglin and Shen, Zhimeng and Huang, Jinghao and Li, Bin and Xue, Xiangyang},
  journal={arXiv preprint arXiv:2306.09682},
  year={2023}
}