Peter Lin
I don't have c++ sample code, but I will keep this issue open for others to answer.
看来TensorRT只支持Resize层用固定的Factor，也就是 downsample_ratio 不支持更改。你可以自己写ONNX导出代码把downsample_ratio固定，然后转TRT
@hust-kevin 你的输入必须FP16/32 要和模型符合，downsample_ratio需要为FP32。看文档
提供的代碼只能換背景，你可以導出綠背景，或自己寫代碼更換背景為視頻。
Looks like this is more a libtorch issue rather than our model issue. Please try run other models to make sure it is not your environment problem.
1. PhotoMatte 的数据是通过绿屏抠像得到的，质量不如 Distinctions，所以后训练 Distinctions。 2. 是。 3. 都可以吧。
Our new method does that. https://peterl1n.github.io/RobustVideoMatting/
Read the paper yet? https://arxiv.org/pdf/2012.07810.pdf
You do need background data https://grail.cs.washington.edu/projects/background-matting-v2/#/datasets
Just download the dataset. Update `data_path.py`, and the training script will composite images on the fly