Detectron.pytorch icon indicating copy to clipboard operation
Detectron.pytorch copied to clipboard

Published 20 hours ago • verified publisher icon roytseng-tw

Adding support for Pascal Voc and custom data

Open adityaarun1 opened this issue 6 years ago • 9 comments

Hi @roytseng-tw ,

Thanks for writing this amazing repo.

I have added (explicit) support for Pascal VOC 2007 and 2012 data set, along with a custom data set. I have also added a basic script to convert XML annotations to COCO json format. You can take a look at the current changes here.

Currently, I am benchmarking the results for VOC 2007 for various settings. I will add them to BENCHMARK.md once they are done.

If you think that these minor additions would add to this repo, I will be happy to open a pull request. If you think that there is something missing/wrong, I can add/fix them.

Thanks, Aditya

System information

  • Operating system: Ubuntu 16.04 LTS
  • CUDA version: 9
  • cuDNN version: 7
  • GPU models (for all devices if they are not all the same): 4x1080Ti
  • python version: 3.6
  • pytorch version: 0.4

adityaarun1 avatar May 24 '18 15:05 adityaarun1

Hi @adityaarun1, It's nice to have VOC dataset support. I'll go check it out.

roytseng-tw avatar May 25 '18 14:05 roytseng-tw

@adityaarun1 Do you train res101-C4 or res101-FPN on voc07 trainval? I can only get Mean AP about 0.72 on voc 07 test.

philokey avatar May 28 '18 03:05 philokey

@philokey I haven't tested it completely yet. With res101-C4 I was getting 0.751 mAP which looked reasonable. I have to run proper benchmarks though for various settings.

Can you provide details for the particular setting for which you got 0.72 mAP?

adityaarun1 avatar May 28 '18 04:05 adityaarun1

@adityaarun1 My settings are pasted in this link. By the way, can you public your setting with 0.751 mAP?

philokey avatar May 28 '18 06:05 philokey

@adityaarun1 Thanks for your work. And i'm confused of that how should i prepare the VOC2007. just put it in the root path or i need to transfer it to the coco format

Hsintao avatar Jun 14 '18 07:06 Hsintao

Hi @Hsintao, data for VOC needs to be in COCO's JSON format. Read more about preparing data here.

P.S.: To answer the previous question, the setting for which I got 0.751 mAP can be found here.

adityaarun1 avatar Jun 14 '18 08:06 adityaarun1

@philokey You need to increase numbers of GPUs so as to increase batch_size

engineer1109 avatar Jul 19 '18 10:07 engineer1109

Hi @Hsintao, data for VOC needs to be in COCO's JSON format. Read more about preparing data here.

P.S.: To answer the previous question, the setting for which I got 0.751 mAP can be found here.

Could you please tell us which step you pick as the final model?

szrlee avatar Oct 09 '18 21:10 szrlee

@szrlee I used 60,000th iteration as my final model.

adityaarun1 avatar Oct 09 '18 23:10 adityaarun1