faster and more informative training plots

[jveitchmichaelis]

Per the discussion here: #1096

• plots generated using OpenCV rather than Matplotlib for speed
• display object confidence and name if available
• class colours are consistent

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Refined image plotting during model testing and training in YOLOv3.

📊 Key Changes

• 🎨 Removed initial image plotting with bounding boxes during the test phase.
• ✨ Added separate image plotting for ground truth and predictions after model output processing during testing.
• 📈 Updated training image plotting to send the image result directly to the TensorBoard writer.
• 🧠 Created a new utility function output_to_target for converting model output to a target format.
• 👁️ Overhauled plot_images function to handle image resizing, improved labeling, and plotting enhancements.

🎯 Purpose & Impact

• 🚀 Enhances the visualization of object detection results by clearly delineating between ground truth and predicted bounding boxes.
• 🛠️ Simplifies the TensorBoard logging code, making integration smoother and more efficient.
• 📏 Helps with interpretation of model performance by introducing a consistent target format, aiding both debugging and analytics.
• 🖼️ Improves the image plotting functionality making it more flexible for various image sizes and richer for analysis with better labeling options.

[glenn-jocher]

@jveitchmichaelis thanks a lot for your contribution, I think this will help everyone tremendously!

[glenn-jocher]

@jveitchmichaelis I did a bit of cleanup in 0ffbf55. Most of the cv2 box plotting can be handled by the existing function used in detect.py:

yolov3/utils/utils.py

Line 862 in 0ffbf55

def plot_one_box(x, img, color=None, label=None, line_thickness=None):

[glenn-jocher]

Uploading images here to use urls in README:

[glenn-jocher]

[glenn-jocher]