YOLO Sleeping Pose Detection
The training was performed in two steps using a common machine learning strategy called transfer learning and fine-tuning.
Initial Training with yolo11s.pt (10 epochs):
Transfer Learning: yolo11s.pt is a pre-trained model, meaning it has already learned to detect a wide variety of objects from a very large dataset (like COCO). Starting with a pre-trained model saves significant training time and computational resources, as the model already possesses general feature-extraction capabilities. Rapid Adaptation: The initial 10 epochs allow the model to quickly adapt its learned features to the specific classes and characteristics of your custom dataset (Object-Detection-1). It begins to adjust its weights to better recognize your unique objects. Second Training with runs/detect/train/weights/best.pt (100 epochs):
Fine-tuning from the Best State:
After the first 10 epochs, the model saves the weights that achieved the best performance on the validation set during that initial run as best.pt. The second training run then continues training from this best.pt file.
Extended Optimization:
This second phase is a more prolonged fine-tuning process. By starting from the best.pt of the previous run, the model doesn't have to re-learn its initial adaptation. Instead, it builds upon the already good performance achieved in the first phase, allowing for more epochs (100 in this case) to further optimize the weights for even better accuracy and generalization on your specific dataset. Iterative Improvement: This approach also enables iterative improvements. You can evaluate the model after an initial short training phase, and if satisfied with the progress, you can continue training for more epochs or experiment with different parameters, always starting from the best performing model achieved so far.
In summary, this two-step process efficiently leverages general knowledge from a pre-trained model and then systematically refines that knowledge to maximize performance on your specific object detection task.