Holes in the object

[egorovivannn]

I am not getting persistent results when I try to retrain your model from scratch on my data (and even fine tuning too). I have data that consists of objects on white background and cycling on 360° and for training I use every 10th frame of a video. The data is not that hard to remove the background from my point of view. But the model is struggling at this data. On the pictures below you can see the original and the result and this seems strange to me, I think I am doing smth wrong. The whole training process doesn't differ from your pipeline in terms of LR, epochs and so on. I use videos from my data as video matte dataset, I use high quality images for Image Matte dataset and some low resolution dataset for segmentation. Overall amount of data is 17k frames. What the problem is and how I could improve the situation? And if anyone managed to retrain the model on another domain, could you share your code, experience of thoughts on my issue?

[buzuyun]

hi，have you solved your problem? i meet the same problem