Inference

[ssy-sheshe]

There is a way of inference by segmenting and tracking Anything in the paper，but I don't find this demo in the repo.Can you please provide it for me?

[siyuanliii]

Thank you for the question. To use the track and segment anything model with MASA, we need to use MASA's distilled detector, you need to add codes here:

masa/masa/models/mot/masa.py

Line 350 in 1d5d42c

else:

, you first add one condition to indicate using masa's own detector, then you call ref_rpn_results_list = self.rpn_head.predict(xxx) and det_results = self.rpn_head. predict_bbox(xxx) to get region proposals from the RPN head and detection bboxes from the RCNN head. For the details, please take a look of this function: predict_bbox(xxx) and rpn_head.predict(xxx) .
Then, those tracked bounding boxes can be tracked and passes to SAM to get their masks.
Please let me know if you have further questions and good luck with your research!