Open tohsakask opened 2 years ago
The gt is converted from binary mask to polygon segmentation. It may have slightly differences, but will not have large impact on performance.
The gt is converted from binary mask polygon. It may have slightly differences, but will not have large impact on performance.
But I found it has a large impact on performance, for LIID, mAP50: 48.4 -> 52.1, for IRnet mAP50: 46.7 -> 50.5.
But I found it has a large impact on performance, for LIID, mAP50: 48.4 -> 52.1, for IRnet mAP50: 46.7 -> 50.5.
Can you explan how to get those results in details?
Maybe you can visualize the gts before and after conversion.
For fully supervised learning, binary mask and polygon segmentation formats in detectron2 have simmilar performances, so I think they should have simmilar results in weakly supervsed learning.
But I found it has a large impact on performance, for LIID, mAP50: 48.4 -> 52.1, for IRnet mAP50: 46.7 -> 50.5.
Can you explan how to get those results in details?
Maybe you can visualize the gts before and after conversion.
For fully supervised learning, binary mask and polygon segmentation formats in detectron2 have simmilar performances, so I think they should have simmilar results in weakly supervsed learning.
I just tested the same detection result with two different GT's based on COCO API.
I will visualize the gts before and after conversion. Maybe I had done something wrong.
Hi~ I found that there is a difference between the gt generated by the code and the official one provided by VOC.