search

davheld / GOTURN

Source code for paper: Learning to Track at 100 FPS with Deep Regression Networks, Held, et al. ECCV 2016
http://davheld.github.io/GOTURN/GOTURN.html
MIT License
882 stars 316 forks source link

Goturn trained on ALOV only #68

Open subhadip3 opened 6 years ago

subhadip3 commented 6 years ago

Hi, I ran a Goturn(Tensorflow) implementation on ALOV dataset only. When I run it on a test set I find all the bounding boxes at the centre of the image. Does this make sense? Since among majority of the search/target(where obj is centred) pairs in ALOV there is not too much change.

anjali1601 commented 6 years ago

@subhadip3 Did you find a solution?

davheld commented 6 years ago

1) My code generates synthetic image "shifts" by cropping the image in different ways, so this bias in ALOV shouldn't cause too much of a problem.

2) Also, at test time the bounding box is typically initialized at the location of the object and then the output prediction is a relative motion, relative to the location of the box in the previous frame. So the box at test time shouldn't necessarily go to the center of the image unless that's where it starts.