Training performance of TExt detection model FCENET degrading on increasing epochs on CTW1500 style custom dataset

[manusid123]

We are trying to train text detection model FCENET on custom dataset which is of the style of CTW1500 dataset with curved text on objects but the object is focused one.

We tried to train FCENET for 500 epochs ona custom dataset of 4k images with some augmentations but we found that pre-trained model is giving better performance than trained models.

Also hmean precision and recall are also decreasing with increasing epochs

Could anyone please suggest whether we are doing something wrong on training side ?

[qiyueliuhuo]

I also seem to encounter such a problem. The more the training effect is, the worse the effect. I am trained on another data set.