Whether soil can be used as a target for detection

[chenfh21]

Hi: I'm curious about soil as an object in the Phenobench plant object detection task. I checked the objects in the MSCOCO dataset (80 & 91), and I understand the soil can be concluded as the 'stuff' category as it is without clear boundaries. I'm confused about how Phenobench determines it as an individual object. And whether transfer learning was used during training (I carefully read your latest paper but couldn't find any information on whether pre-trained weights were used in the 300 iterations).

In addition, I use the YOLOV7 default settings when conducting experiments on the network I proposed, and all other indicators performed well except for the increase in val/obj-loss. I would like to know if you have also encountered such a problem. The reason given by the community is the inconsistency caused by using computeLossOTA during training and computeLoss during validation (ref: https://github.com/WongKinYiu/yolov7/issues/326).