Open hassanaliemon opened 2 years ago
When the height-width-ratio of object is larger than 4x height-width-ratio of anchor, the object will never get positive samples for training. For this case, you have to add more anchors to fit the long objects.
@WongKinYiu @AlexeyAB how can we add anchor? is there any method like alexyab darknet has ? in which we calculate our custom anchor based on the dataset?
@akashAD98 you can use google to get some scripts that will help you to get your anchors, however when you are working with your won data I think it will just come to you with time.
@WongKinYiu thanks for the reply. I visualized the targets and after changing anchors no target is missing now. However, I am getting that long/big object problem still! I have come to conclusion that Small object detection is excellent but medium to long/big objects are very poor in detection.
As I have told, I have visualized targets right before it calculates loss and right after 'targets' variable comes from 'self.build_targets' function. Is there anything I am missing?
@hassanaliemon it would be great if you can do this script as open source for community
Hi, I am training yolov7 to detects objects and it detects small & medium objects accurately. But the problem arises when long objects come, it performs poor in single long objects (along with small & medium objects). However, when several long objects are situated closely, model is no where near to accurate detection.
Previously it was not detecting long objects at all, then I turned augment to false. Now, it detects long objects but performance is very poor.
What can be the problem? Thanks for this amazing repo :)