Target score drops below 0 very fast

[AlexTS1980]

I'm f'tuning MDNet to my data. Only the first frame in the video is labelled with a bbox. After 8-10 frames the target score drop below 0. I played with learning rates/multipliers/trainable layers, but no success so far. Anything I could do or I need to label a larger number of frames in the dataset?

[hyseob]

Hi @AlexTS1980,
You can also try reducing the update interval (tracking/option.yaml:long_interval) or increase training iterations (maxiter_update), although I recommend to keep the default options for general performance.
MDNet is not always perfect and may fail if your target undergoes severe occlusion or rapid deformation.

[AlexTS1980]

Thanks. Not really the occlusion/deformation, but very high degree of similarity, like this:

[ghost]

Hi @HyeonseobNam
how you set the number of the long_interval and iterations ?
why iterations is 10?
thanks