@Jingkang50 ,Hi, according to the work you have done, I have created a small dataset specifically for railway scenarios. It only includes two classes of objects: pedestrians and railways. It also includes two types of relationships: "on" and "beside." After training for 100 epochs multiple times, I noticed that the results are consistently zero. I obtained this result when using the predict file to infer an image. As you can see, the PSGTR model I trained predicted the entire image as "person."Can you give me some advice?
@Jingkang50 ,Hi, according to the work you have done, I have created a small dataset specifically for railway scenarios. It only includes two classes of objects: pedestrians and railways. It also includes two types of relationships: "on" and "beside." After training for 100 epochs multiple times, I noticed that the results are consistently zero. I obtained this result when using the predict file to infer an image. As you can see, the PSGTR model I trained predicted the entire image as "person."Can you give me some advice?![截屏2023-06-20 16 33 22](https://github.com/Jingkang50/OpenPSG/assets/82081034/23446e9d-a311-4ea8-9fa7-33713f509726)