Multiple stacks of line along side actual line, Custom training on MLSD-large

lhwcv / mlsd_pytorch

Pytorch implementation of "M-LSD: Towards Light-weight and Real-time Line Segment Detection"

Apache License 2.0

189 stars 37 forks source link

Multiple stacks of line along side actual line, Custom training on MLSD-large #23

Open sakshamprakash01 opened 1 year ago

sakshamprakash01 commented 1 year ago

Training was done using the following parameters-

It still had High Recall (0.9) But Low Precision (0.2) and sAP of around 35.6; F-score suffers due to low precision. Loss in early 20s.

How do I suppress the noisy lines and enhance the detections? Maybe some sort of NMS-equivalent is present?

xu-peng-7 commented 1 year ago

Hello, are you also researching line segment detection based on drawings? I also started to do the whole aspect recently, can you communicate with me?

mlsd I think the error may be downsampling or nms, I have no similar problems in using lsd。

Lyn4444 commented 11 months ago

i think batchsize is small, my batchsize is 32, sAP is 42