search

ViTAE-Transformer / SAMRS

The official repo for [NeurIPS'23] "SAMRS: Scaling-up Remote Sensing Segmentation Dataset with Segment Anything Model"
286 stars 14 forks source link

Can I use this code to get other dataset masks #24

Open lsj1111 opened 4 months ago

lsj1111 commented 4 months ago

it's a good job，and I saw in your paper that you obtained the segmentation results of DOTA 2.0 version. Can I use this code to obtain the segmentation results of DOTA 1.0? look forward to you reply, thank you very much

DotWang commented 4 months ago

@lsj1111 Of course, we have released the code. As I know, DOTA-v1.0 is actually a subset of DOTA-v2.0 and they have the same label format, thus it is easy to use our code for performing the label conversion.

DotWang commented 4 months ago

@lsj1111 Even if for other detection datasets, our code is still available, but you may need to manually change the code of reading detection labels according to corresponding formats.

lsj1111 commented 4 months ago

@lsj1111 Even if for other detection datasets, our code is still available, but you may need to manually change the code of reading detection labels according to corresponding formats.

OK, I got it ,thank you very much

lsj1111 commented 4 months ago

@lsj1111 Even if for other detection datasets, our code is still available, but you may need to manually change the code of reading detection labels according to corresponding formats.

When I changed the data set to dota1.0, there were more than 20,000 pictures in total, but the output segmentation results were only more than 13,000. What is the problem?According to the printed information, it can be seen that there is indeed a problem during the segmentation process： dota dataset contains 21046 images

DotWang commented 4 months ago

@lsj1111 The segmentaation map cannot be produced if there is no box, some samples in DOTA 1.0 have not been annotated.

lsj1111 commented 4 months ago

@lsj1111 The segmentaation map cannot be produced if there is no box, some samples in DOTA 1.0 have not been annotated.

yes,you are right,i got a new problem, for the same picture, I used DOTA_devkit to crop the picture. For detection tasks, the annotation when cropping is usually Rbox, but in your paper I found that using hbox as a prompt when predicting the mask will be better, but these two annotations The Ground Truth results obtained by different methods are different. e64f775dab05dd9b1abe13e3b1586f4 The left side is the result obtained by Hbox annotation, and the right side is the result obtained by Rbox. Do you have a better solution?

DotWang commented 4 months ago

@lsj1111 Firstly, for R-Box, you should use RH-Box as the prompt

In addition, even for a same picture, the crop results may be different if using different types of boxes, some boxes may be removed during the cropping.