How does offset regression really work?

[mridvankeles]

I couldnt really understand the idea and the importance of the junctions and heatmap,offset regressions in this repo.

at the encoder.py ; xint, yint = junctions[:,0].long(), junctions[:,1].long() off_x = junctions[:,0] - xint.float()-0.5 off_y = junctions[:,1] - yint.float()-0.5

this part doing the offset convertion from junction x and y points i see that. And at the inference phase, we use those offsets to get score for nms postprocess. I couldnt find much more details neither on the paper nor on the internet.

ill be preciated if anyone helps. thx

[SarahwXU]

The separate heatmap and offset ensure the accuracy of junction detection for predicting the junction location on the 1/4 resolution feature map. You may refer to papers like ''Holistically-Attracted Wireframe Parsing'' and ''End-to-End Wireframe Parsing".

[mridvankeles]

Now i understand it. Thanks for ur work and reply.