Bug fix mAP0.5-0.95

[lebedevdes]

Two changes provided

1. Added limit on the maximum number of detections for each image likewise pycocotools
2. Rework process_batch function

Changes #2 solved issue #4251
I also independently encountered the problem described in issue #4251 that the values for the same thresholds do not match when changing the limits in the torch.linspace function.
These changes solve this problem.

Currently during validation yolov5x.pt model the following results were obtained:
from yolov5 validation
Class Images Labels P R [email protected] [email protected]:.95: 100%|██████████| 157/157 [01:07<00:00, 2.33it/s]
all 5000 36335 0.743 0.626 0.682 0.506
from pycocotools
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.505
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.685

These results are very close, although not completely pass the competition issue #2258.
I think it's problem with false positive bboxes matched ignored criteria, but this is not actual for custom datasets and does not require an additional solution.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhancements to type conversions and matching criteria in YOLOv5 validation code.

📊 Key Changes

• Updated numpy type conversions from .astype('int32') and .astype(np.int16) to .astype(int).
• Refactored matching logic in process_batch to iterate over different IoU (Intersection over Union) thresholds.
• Changed tensor indexing to use .astype(int) in val.py, ensuring compatibility across data types.

🎯 Purpose & Impact

• The type conversion change ensures more general integer representation for better compatibility across platforms.
• The improved matching logic allows for more precise performance evaluation at different IoU thresholds, likely leading to more accurate model validation results.
• These updates may provide a slight performance increase and ensure more consistent behavior in different environments, benefiting both developers working with the code and users relying on the model's accuracy.

[github-actions]

👋 Hello @lebedevdes, thank you for submitting a YOLOv5 🚀 PR! To allow your work to be integrated as seamlessly as possible, we advise you to:

• ✅ Verify your PR is up-to-date with upstream/master. If your PR is behind upstream/master an automatic GitHub Actions merge may be attempted by writing /rebase in a new comment, or by running the following code, replacing 'feature' with the name of your local branch:

git remote add upstream https://github.com/ultralytics/yolov5.git
git fetch upstream
# git checkout feature  # <--- replace 'feature' with local branch name
git merge upstream/master
git push -u origin -f

• ✅ Verify all Continuous Integration (CI) checks are passing.
• ✅ Reduce changes to the absolute minimum required for your bug fix or feature addition. "It is not daily increase but daily decrease, hack away the unessential. The closer to the source, the less wastage there is." -Bruce Lee

[glenn-jocher]

@lebedevdes resolved conflicts and updated PR. I will take a look at this, I don't know why it's been left open so long.

[glenn-jocher]

@lebedevdes it seems like the pred = pred[:100] line is reducing the pycocotools mAP. If I remove it pycocotools mAP remains unchanged while YOLOv5 mAP slightly reduces as well, which is probably the preferred solution as it still represents a big improvement over current.

[glenn-jocher]

@lebedevdes PR is merged. Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐

[lebedevdes]

@glenn-jocher Glad to help, thanks for your great project)

[glenn-jocher]

@lebedevdes thanks for your kind words, but the real credit goes to the YOLO community and the Ultralytics team. We greatly appreciate your contributions! If you have any more ideas or feedback, feel free to share them with us.