search

yangxue0827 / RotationDetection

This is a tensorflow-based rotation detection benchmark, also called AlphaRotate.
https://rotationdetection.readthedocs.io/
Apache License 2.0
1.08k stars 182 forks source link

problem with dataset setting…… #23

Closed biBottle closed 3 years ago

biBottle commented 3 years ago

sorry， still me……maybe I am too stupid, sorry to bring trouble to you..... the problem yesterday does not solve yet. Now the result imgs are like this(I change the self.cfgs.VIS_SCORE so it can show some boxes but obviously not correct). the dataset and command I use are like following. can you help me to check is there anything wrong TAT 2021-06-17 09-47-55屏幕截图 2021-06-17 09-39-18屏幕截图 2021-06-17 09-39-39屏幕截图 2021-06-17 09-39-59屏幕截图 2021-06-17 09-40-17屏幕截图 2021-06-17 09-50-41屏幕截图 2021-06-17 09-58-58屏幕截图

hong123123 commented 3 years ago

Maybe there's no problem with your pipeline, just the performance of the model is unsatisfying.

So I recommend you to test with images in training set.

yangxue0827 commented 3 years ago

Obviously, you are still loading the pre-trained weights. Please confirm whether the VERSION in cfge.py and the trained weights you downloaded match. @biBottle

yangxue0827 commented 3 years ago

image

biBottle commented 3 years ago

I do as your advice and find there is one issue may also help....

In the checkpoint in r2cnn of dota1.0(the folder name the trained model is FPN_Res50D_DOTA_1x_20201031) ,it writes [model_checkpoint_path: "DOTA_486003model.ckpt"], when I change it to [model_checkpoint_path: "DOTA_378003model.ckpt"](the name of other files in the folder), then it works.

Thanks for the help!