simplify the way to load annotation file

[developer0hye]

simplify the way to load annotation file

@glenn-jocher

How about this way in terms of readability?

The below code is simple example code.

import numpy as np

gt_file = np.array([[0, 0.5, 0.5, 0.5, 0.5],
                   [1, 0.5, 0.5, 0.5, 0.5]]).reshape(-1, 5)

np.savetxt("gt_file.txt", gt_file)

gt_file = np.loadtxt("gt_file.txt", dtype=np.float32).reshape(-1, 5)

print(gt_file)

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Simplified label loading process in YOLOv5 datasets.

📊 Key Changes

• Removed complex label parsing logic with multiple condition checks.
• Replaced with a streamlined np.loadtxt for direct reading and reshaping of label files.

🎯 Purpose & Impact

• Purpose: To streamline the process of loading labels for images, making the codebase simpler and potentially faster.
• Impact: Users should experience more efficient label loading with fewer opportunities for errors during the label verification process. 🚀

[developer0hye]

It is very slow...

[glenn-jocher]

@developer0hye yes reading label text files is pretty slow. Label text files can be of two formats:

Box labels

class x y w h
class x y w h
...

Segment labels (can be any number of segment points > 4, different rows can be different length)

class xy0 xy1 xy2 xy3 xy4  # minimum 9 columns
class xy0 xy1 xy2 xy3 xy4 xy5 xy6 ...  # no maximum columns
...

[developer0hye]

@glenn-jocher

Ah, okay. Original one is better than this!

[glenn-jocher]

@developer0hye well I don't know, if there's a faster way to load and parse the data depending on the column count then that would be great.