search

XinzeLee / PolygonObjectDetection

This repository is based on Ultralytics/yolov5, with adjustments to enable polygon prediction boxes.
352 stars 95 forks source link

How much data do i need for custom training? #33

Open Bigdata92 opened 2 years ago

Bigdata92 commented 2 years ago

Currently, I am custom-training Polygon-Yolov5, but it is not trained properly even after training 40 trains and 10 vals like the UCAS50 dataset

If you know what I did wrong please tell me

1 2

futureflsl commented 2 years ago

Currently, I am custom-training Polygon-Yolov5, but it is not trained properly even after training 200 epoch,P,R ,map is zero all the time.this project seems to has some problem.

If you know what I did wrong please tell me

XinzeLee commented 2 years ago

@Bigdata92 I cannot tell how many data you used. But if the amount of training data is less than 1000 for each class, please add more. @futureflsl I trained on a relatively large dataset (around 1000 for each class) for training, and test the trained model on a small sampled dataset UCAS50. It does not necessarily mean that you can get good results from training on such small datasets.

Aun0124 commented 1 year ago

@Bigdata92 Hi, can i have the codes or idea on your works. I'm doing similar stuff. Thank you

lemon1022 commented 1 year ago

@Bigdata92 Did you solve it?

FrankTHALES commented 1 year ago

Hi @Bigdata92 , we have the same issue and we have the same kind of images in our dataset... did you progress to the resolution of your problem, we are very interested...

lemon1022 commented 1 year ago

@FrankTHALES I have solved this problem now, mainly because the data set is too small, and another reason is that the features obtained by the model at the initial stage of training are not obvious and cannot be recognized. The model I am training now is in the first two rounds. P, R, mAP are still 0, but it will be normal by the third round.