I need to perform transfer learning on a class of a pre-trained model using pytorch oriented to object detection.
Basically what I want to do is based on some detections that I have done manually on a series of cars, perform transfer learning to improve the output of the detection model.
To date I have not found information related to this aspect, so I do not know very well how to perform this task.
The transfer learning technique you are looking for in this case is called fine-tuning. Basically you take a network for object detection (a CNN architecture for instance, from
torchvision.models), and you “unfreeze” some of its layers, and train only those. You usually want to unfreeze the last layers, as the first ones are used to extract more general features. The more images you have in your own labeled data, the more layers you should “unfreeze”. Probably a good starting point is this, but if you want more just Google “fine tune object detection pytorch”.