Open artist-duan opened 3 years ago
Hi js-duan. Sorry for the late reply. Yes, bbox does converge (in my custom dataset and also on coco-dataset). For the inference time, I've implemented cc_fast-nms (please get the latest pull), which is much faster. Also, while comparing speed with the original implementation keep in mind whether the original implementation is considering the post processing while calculation the speed or not. As in this version of code, post processing is in detection.py file, which multiplies proto-coef and also crop mask, which will always be included in saved_model.pb, so this implementation always include post-processing while infering, while the original does not.
Does the bbox regression branch converge? I train yolact on my dataset, but the regression is not work! and the inference time is slow, is twice as much as in the paper, Do u encountered the same problem?