[Feature] I've added support for YoloTxt Dataset Format

[NWalker4483]

What is the feature?

Why?

1. I've been working with a lot of synthetic data and find the txt format makes it easier to experiment with my tools and it's straightforward
2. A lot of people already use this format for their computer vision projects

How it works

The dataset follows this structure:

dataset_root/
├── train/
│   ├── images/
│   │   ├── image1.jpg
│   │   └── image2.jpg
│   └── labels/
│       ├── image1.txt
│       └── image2.txt
└── val/
    ├── images/
    │   ├── image3.jpg
    │   └── image4.jpg
    └── labels/
        ├── image3.txt
        └── image4.txt

Label Format

Each .txt file contains one annotation per line in the format:
{class_id} {x_center} {y_center} {width} {height} {keypoint_x1} {keypoint_y1} {visible1} {keypoint_x2} {keypoint_y2} {visible2} ...

Where:

• All values are normalized to [0, 1]
• visible is 0 for not labeled, 1 for labeled but not visible, 2 for labeled and visible

Usage

train_dataloader = dict(
    ...
    dataset=dict(
        type='YOLOPoseDataset',
        data_root=data_root,
        data_mode=data_mode,
        data_prefix=dict(img='train'),
        pipeline=train_pipeline,
        metainfo=dict(from_file=metainfo_file)
    ),
)

Is this considered within scope for a PR?

Any other context?

No response

[NWalker4483]

To be clear, this is something I'm implemented already for my project. I'm just curious if it's within scope to be merged by the repo