Best/ good pre-trained models for extracting feature from image

vision
1

Hello guys. As the title suggests, I am interested in extracting features from an egomotion image (an image taken from a car’s front dashcam). At the moment, I am using Resnet-50 but I don’t know whether this is the best one for this purpose. Honestly, I don’t understand how you would choose the appropriate neural network for this purpose.

Some more context:
I am building a machine learning model for driving-related purposes. I want to predict the trajectory of objects that my car sees. For now, my plan is to extract features from the image at every time step and input those to an RNN encoder. With that in mind, what neural network is most likely to extract appropriate features. After extracting the features, I push them through a single fully-connected layer before going to the RNN encoder stage.

2

By Resnet-50 do you mean an imagenet pre-trained one? If you have enough data, one option to get some good features would be to train a network (like Resnet or VGG) from scratch using self-supervised approaches like rotation prediction. This will help learn some features specific to your data.

3

Always better to init with pretrained weights, even with own data