Using a custom dataset in the BOP format

[3bsamad]

I am trying to train the network using a custom dataset. I've read I need to have it organized in the BOP format, which I've (almost) successfully done. My data is in .h5 format and I was able to extract the data in a similar fashion.

My only problem is the file scene_gt_info.json which contains info like bbox_obj and bbox_visib, which, according to the website, needs depth information to calculate the visible parts/ pixels etc. I wanted to train the RGB only GDRNet, since I generated only RGB data. It seems the GDRNet itself does not necessarily need the depth info, but then how can I generate a file similar to scene_gt_info.json to feed to the network?

[wangg12]

If you have CAD models, you can get bbox_obj and bbox_visib from rendered masks.

[3bsamad]

I have CAD models, I'll look into this. Does this also mean that the network doesn't need/expect all the information the scene_gt_info.json, and that I can just provide these two?

[Lokesh-26]

Hello @wangg12,
I have the scene_gt_info.json of my own dataset. How do I obtain the the test_bboxes for a custom dataset?

[wangg12]

@Lokesh-26 Just provide the detection results either from a detector or from other ways. Make sure the format in test_bboxes adheres to the code (refer to the code or the existing test_bboxes examples in README).