search

dbolya / yolact

A simple, fully convolutional model for real-time instance segmentation.
MIT License
5.01k stars 1.32k forks source link

Anchor box viz #154

Open abhigoku10 opened 5 years ago

abhigoku10 commented 5 years ago

@dbolya can i visualize the anchor boxes on the test image during inference like in this https://github.com/matterport/Mask_RCNN since i need to debug the detection area

dbolya commented 5 years ago

Not explicitly, but you can try multiplying all the predicted bbox loc regressors in the prediction head by 0. This will only pass unregressed anchors through the network.

After this line: https://github.com/dbolya/yolact/blob/09f2bc652c6c01aaf7b99abbb96860e919934aa9/yolact.py#L213 Set bbox = bbox * 0

You might have to fiddle with the NMS parameters (or not do NMS at all) to get all the boxes to show up.