search

weekenddeeplearning / TrackNet

Heatmap based high speed tiny sport objects tracking based on TrackNet
GNU General Public License v3.0
43 stars 9 forks source link

Inference result was a total black map #1

Open ztianlin opened 3 years ago

ztianlin commented 3 years ago

Really appreciated for this excellent work. I have tried to use a lighter backbone to train on my own dataset. The loss converged quickly, however, after the first epoch the output map was totally in black (all pixel values were 0). I thought there was an imbalance between negative samples (black pixels in the heatmap annotation) and positives. Can you pull me out of this situation? Many thanks.

weekenddeeplearning commented 3 years ago

@ztianlin - Sorry about getting back late. I had put together this repo by contacting the authors and using their repo below. We would have to go through their train module to understand what's going on. Did you verify the annotation heatmaps and label csv before the training?

[https://nol.cs.nctu.edu.tw:234/open-source/TrackNet/tree/master]

ztianlin commented 3 years ago

Well, It seems that I implemented the network badly. In your keras implementation, the normalization method is layer normalization instead of batch normalization, right?

masouduut94 commented 2 years ago

So any luck with the training? Every time I run training, I get errors during epochs 1-7. @ztianlin