Closed newforrestgump001 closed 2 years ago
Yes, but you need a neural network that does not have some size requirements (e.g. a fully connected layer without some global pooling before it, or an input_rgb_image_sized
layer).
If that is the case, then you can infer any size you want (as long as it fits on memory). That being said, if you use varying sized images, you can't infer them by batches, you have to do it one by one. Moreover, the network was probably trained with a particular size in mind (plus some data augmentation) so if your images differ too much from the original training set, you'll experience a drop in performance.
@arrufat Sincerely thank for your quick reply. Does dlib dnn module support 3d-cnn, for example, used for video classification?
No, it doesn't.
@arrufat I got it, Thank you ~~
@arrufat Sorry to interrupt you again. It is possible to replace svm classifier with one-class svm in self supervised demo (examples/dnn_self_supervised_learning_ex.cpp) in dlib. If possible, I want to try anomaly detection.
I guess so, right away, I can't see why you couldn't do it.
Got it. Thank you!
I have vary-size images, it is possible to classify them without resizing them to same size. Thank you a lot!