---
language:
- en
tags:
- robotics
---



## EgoPAT3Dv2

### Dataset introduction
There are **11 scenes** contained in the EgoPAT3Dv2 dataset, corresponding to folders 1 through 11. Each scene folder contains 2 to 6 video folders, and each video folder contains an **RGB** folder, a **depth** folder, a **point cloud** folder and a **transformation matrices** folder. (Please ignore other folders or files inside the zip file.) The annotations (ground truth) and transformation matrices (the same as the transformation matrices above) are included in the annotation_transformation.hdf5 file. We use HDF5 to organize the dataset in the experiment, and the dataloader in the GitHub repo is also written correspondingly.


### Dataset folder hierarchy
```bash
Dataset/
    ├── 1	# scene 1
        ├── 1.1.zip -> 1.1		# video 1 in scene 1
            ├── d2rgb						# depth files
            ├── color						# rgb files
            ├── pointcloud			# point cloud files
            └── transformation	# transformation matrices
        ├── 1.2.zip -> 1.2 	# share the same structure as 1.1
        ├── ...
        └── 1.4.zip -> 1.4
    ├── 2/ # all scene/video directories share the same structure as above
        └── ...
    .
    .
    .
    
    └── 11
```

## Construct HDF5 dataset file

Since 50GB is the hard limit for single file size in huggingface, please use [make_RGB_dataset.py](https://huggingface.co/datasets/ai4ce/EgoPAT3Dv2/blob/main/make_RGB_dataset.py) to construct the HDF5 file on your own.

1. Download all zipped files. Unzip them and keep RGB("color" in the folder) folder in each video folder only. 
2. Run `make_RGB_dataset.py` after step 1.