search

bertinetto / siamese-fc

Arbitrary object tracking at 50-100 FPS with Fully Convolutional Siamese networks.
http://www.robots.ox.ac.uk/~luca/siamese-fc.html
MIT License
628 stars 224 forks source link

net-train graph issue #15

Closed boldjoel closed 7 years ago

boldjoel commented 7 years ago

Hey, thanks a lot for your work. And I follow your instruction and try to produce my own model for tracking. However, the train graph looks really weird since the errmax is almost 0 from the start of training net-train.pdf. I train this model with around 3500 frames from video. Just wonder if you have any clue for the training.

bertinetto commented 7 years ago

Hi, You shouldn't worry about errmax when you are not training with negative pairs (which is the default).

The problem here is that it the network it is overfitting really badly, as you can see from the increasing validation error on the loss.

We observed that a bit of overfitting doesn't necessarily means worse performance in tracking. I think this is because the problem the network is trying to minimize (matching a pair of images) is not strictly the same of tracking, where ideally we would like to minimize the overall IOU. However, the case you presented is overfitting really badly, you should take care of that. I feel the culprit is the small amount of training frames you are using.

Hope this helps.