Only use annotated images to create dataset in autosplit()

[kinoute]

After I read your message to get the best results from Yolov5 in #2313 (comment) where you talk about background images and also after our discussion regarding autosplit() and annotated files in #2228, I decided to add two functionalities in autosplit():

• The first one is the ability to create a dataset/splits only with images that have an annotation file, i.e a .txt file, associated to it. As we talked about this, the absence of a txt file could mean two things:

  • either the image wasn't yet labelled by someone,
  • either there is no object to detect.

  When it's easy to create small datasets, when you have to create datasets with thousands of images (and more coming), it's hard to track where you at and you don't want to wait to have all of them annotated before starting to train. Which means some images would lack txt files and annotations, resulting in label inconsistency as you say in Hunt for the highest MAP #2313. By adding the annotated_only argument to the function, people could create, if they want to, datasets/splits only with images that were labelled, for sure.

• The second functionality is the ability to fill our dataset with background images automatically. Basically, we provide a path in the bg_imgs_path argument and some images from this path/folder will be picked and added to the splits. The number, or more precisely, the ratio of background images can be configured with the bg_imgs_ratio. I followed your advices on the other issue and limited it to a float between 0 and 0.1 (0% and 10%). This ratio will be used to calculate how many background images should be added in each split. If my training split has 1000 images and I set up bg_imgs_ratio to 0.1, then, if I have enough background images to fill the demand, around 100 background images will be added to the training split.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Updated YOLOv5 Dockerfile, README, and various performance improvements.

📊 Key Changes

• Docker base image updated to a newer PyTorch version.
• Installation of Python packages now includes a notebook and uses a cache.
• Environment variable HOME set for Docker.
• Simplified PyTorch Hub models loading code.
• Argoverse-HD dataset support added with scripts for preparation.
• Default pretrained weights usage is now consistent across model types.
• Performance improvements in data loading functions.

🎯 Purpose & Impact

• 🚀 Better base environment due to newer Docker image.
• ⏱ Reduced Python package installation time via pip caching.
• 🛠 Simplified usage for users loading models with PyTorch Hub.
• 📈 Enhanced dataset support enables users to work with Argoverse-HD, expanding the range of applicable real-world scenarios.
• 🤖 Ensures pretrained models are now the default, providing better out-of-the-box performance.
• ✅ General optimizations lead to faster data preprocessing, benefiting all users in terms of efficiency.