Support for UltralyticsParser.

[kkeroo]

Purpose

A lot of customers are using YOLOv8 or Ultralytics format for datasets. We need a parser for that.

Specification

YOLOv8 dataset format structure is like:

dataset_dir/
        ├── train/
        │   ├── images/
        │   │   ├── img1.jpg
        │   │   ├── img2.jpg
        │   │   └── ...
        │   ├── labels/
        │   │   ├── img1.txt
        │   │   ├── img2.txt
        │   │   └── ...
        ├── val/
        │   ├── images/
        │   │   ├── img1.jpg
        │   │   ├── img2.jpg
        │   │   └── ...
        │   ├── labels/
        │   │   ├── img1.txt
        │   │   ├── img2.txt
        │   │   └── ...
        ├── test/
        │   ├── images/
        │   │   ├── img1.jpg
        │   │   ├── img2.jpg
        │   │   └── ...
        │   ├── labels/
        │   │   ├── img1.txt
        │   │   ├── img2.txt
        │   │   └── ...
        └── *.yaml

while Ultralytics (Hub) format is like:

dataset_dir/
        ├── images/
        │   ├── train/
        │   │   ├── img1.jpg
        │   │   ├── img2.jpg
        │   │   └── ...
        │   ├── val/
        │   └── test/
        ├── labels/
        │   ├── train/
        │   │   ├── img1.txt
        │   │   ├── img2.txt
        │   │   └── ...
        │   ├── val/
        │   └── test/
        └── *.yaml

For example you can download coco8 from Ultralytics Github to get the Ultralytics format, or download any dataset from Roboflow Universe in YOLOv8 format.

This PR introduces all-in-one Ultralytics parser for parsing both formats and following tasks:

• Object Detection
• Instance Segmentation
• Keypoints Detection

How we figure out the task? Object Detection is simple because all 3 tasks have at least 5 elements in each row of an annotation file. All 3 tasks have class_ix, x, y, w, h while the instance segmentation and keypoints have additional ones. If the dataset is keypoints-based then we look into the yaml file that is present in all ultralytics/yolov8 datasets and look for field kpt_shape which tells us the number and dimensionality of the keypoints (can be 2D or 3D). If it is 2D then we add visibility (3rd dim) to 2 (full visible). 3D kpts have 3rd dim already Literal[0,1,2]. So:

• Object detection: len(elements) == 5
• Keypoints: len(elements) > 5 and kpt_shape present in yaml file
• Instance seg: else

Dependencies & Potential Impact

None / not applicable

Deployment Plan

None / not applicable

Testing & Validation

Tested on:

• Coco8 (from ultralytics gh)
• Coco8-pose (from ultralytics gh)
• Tiger-pose (from ultralytics gh)
• Coco8-seg ((from ultralytics gh)
• crack-seg (from ultralytics gh)
• Padel kpts dataset (from Roboflow)
• Fire and Smoke instance seg. (from Roboflow)

[codecov]

Codecov Report

❌ Patch coverage is 91.36691% with 12 lines in your changes missing coverage. Please review.
✅ Project coverage is 95.30%. Comparing base (49a14a0) to head (5161e4d).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
luxonis_ml/data/parsers/yolov8_parser.py	91.17%	12 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff            @@
##             main     #367    +/-   ##
========================================
  Coverage   95.29%   95.30%            
========================================
  Files         104      105     +1     
  Lines        6401     6540   +139     
========================================
+ Hits         6100     6233   +133     
- Misses        301      307     +6     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[klemen1999]

Inital integration looks good, left some comments.
Other comments:

• Add tests
• Add parser to the README and docs.luxonis after merge

[kkeroo]

Tests are added (3 datasets for obj. det, kpt. det., and inst. seg.), README updated, I moved the Ultralytics format up, right after COCO format because its probably very popular