search

jwwangchn / AI-TOD

Official code for "Tiny Object Detection in Aerial Images".
MIT License
193 stars 21 forks source link

About the annotation files of AI-TOD #15

Closed haotianll closed 2 years ago

haotianll commented 2 years ago

Hi,

I follow the orignal instruction and aitodtoolkit and I get some annotation files, but I find the numbers of instances are not consistent as reported in AI-TOD paper. So I want to know if there are any processed annotation files that can be used directly.

I listed the results as following: image

My results:

category_id | aitod_train.json | aitod_val.json | aitod_trainval.json | aitod_test.json -- | -- | -- | -- | -- 1 | 623 | 170 | 793 | 745 2 | 512 | 140 | 652 | 689 3 | 5269 | 2477 | 7746 | 5860 4 | 13539 | 3791 | 17330 | 17633 5 | 293 | 34 | 327 | 292 6 | 248051 | 59906 | 307957 | 306678 7 | 14126 | 3841 | 17967 | 15443 8 | 176 | 67 | 243 | 290 total | 282589 | 70426 | 353015 | 347630
Chasel-Tsui commented 2 years ago

Hi,

I follow the orignal instruction and aitodtoolkit and I get some annotation files, but I find the numbers of instances are not consistent as reported in AI-TOD paper. So I want to know if there are any processed annotation files that can be used directly.

I listed the results as following: image

My results:

category_id aitod_train.json aitod_val.json aitod_trainval.json aitod_test.json 1 623 170 793 745 2 512 140 652 689 3 5269 2477 7746 5860 4 13539 3791 17330 17633 5 293 34 327 292 6 248051 59906 307957 306678 7 14126 3841 17967 15443 8 176 67 243 290 total 282589 70426 353015 347630

Hi, Thanks for pointing it out. I have checked it. The difference may result from the fact that we have provided two versions of AI-TOD in this repo, the statistical information of the first version (the whole AI-TOD was publicly available before 2022/06) is the same as that in the paper. While due to a license conflict with xView, we then disable the public download link of AI-TOD, you may need to generate the AI-TOD from the provided code, obtaining the second version of AI-TOD, your provided result is the statistical information of the second version. The difference may result from the difference in the code of dataset generation. We have also checked that the instances in the two versions only differ marginally which will not have an obvious impact on the final detection performance. If you want to download the previous public version, please provide an e-mail address.

haotianll commented 2 years ago

Thank you so much for your reply! My email address is haotianliu001@hotmail.com .