Can transfer learning approach from 1st example not be used for object detection?

Question

What is the difference between transfer learning using the examples shown below?

1. Image classification - Transfer learning and fine-tuning using pre-trained model (MobileNet V2 model) https://www.tensorflow.org/tutorials/images/transfer_learning#create_the_base_model_from_the_pre-trained_convnets

2. Object detection - See section Create model and restore weights for all but last layer (ssd_resnet50 model) - https://github.com/tensorflow/models/blob/master/research/object_detection/colab_tutorials/eager_few_shot_od_training_tf2_colab.ipynb

Can transfer learning approach from 1st example not be used for object detection?

Answer 1

The difference between these two links is that in the first one, your model is all the pre-trained weights of the MobileNet v2 architecture + a densely-connected prediction layer, which is typically used for image classification. In the second link, that object detection model is the pre-trained weights of the resnet50 model, and the last layer is an SSD layer (single-shot detection, link to original paper here: https://arxiv.org/abs/1512.02325?context=cs). SSD is trained to inference at runtime and classify images + produce bounding boxes around the subjects it is trained to detection.

tldr, the only difference between the two is the last layer attached to the transferred architecture. You can use the resnet architecture for classification by using a dense layer, or for object detection using an SSD layer. Same with mobilenet, vgg, etc.