Custom LSDNet model + Useage on different datasets?

[sakshamprakash01]

The LSDNet model trained by @ZeronSix is made on the wireframe dataset which is based on real-life images.
However, my use case is regarding using it on P&IDs to detect lines (horizontal, vertical and diagonal) in a robust way. These images are mostly black lines/characters on a white background which distinct lines.

A few plots of detections by running the pre-trained model on my images are attached below. The following observations were made-

1. Image size has a huge impact on the detections
2. At places edges of close numbers were detected as false positives.
3. There are true-negatives at places which were unexplainable.

Could you please help with answering the following questions-

1. What is the ideal image size to be used for inferences with the pretrained model?
2. How can one train a custom LSD Net model / do transfer learning on top of the already trained model to adapt to my images?
3. What could be the reason for missed detections in the below images?

[skraina]

@ZeronSix @sakshamprakash01 I'm primarily interested in knowing how can I use this network to train on images which have mixed-type graphics. The dataset which has been trained for this has only one graphic element, that is, line so I'm curious to know what will happen when diverse types of graphics elements such as circles, curves, geometric shapes consisting of small lines, etc. will also be present in the image.

[levTepl]

@sakshamprakash01 , sorry for long reply - was busy. Answering your questions.

What is the ideal image size to be used for inferences with the pretrained model?

Input image is resized to fixed size before the net.
So what is important is typical line size measured not in "pixel size", but in "pixel size" / "image size".

Moreover, the "correct" line size in "pixel size" / "image size" should match those of Wireframe dataset the net was trained on. For your case of custom data - I'd recommend to conduct some experiments in search for optimal parameters.

How can one train a custom LSD Net model / do transfer learning on top of the already trained model to adapt to my images?

If your question is "do we plan to release the training code?" - the answer is "not soon", because it depends on some of ours proprietary data engine and now we have no resources to transfer it to complete opensource. Maybe we will do it someday, when we will have some time.

Anyway, you can rewrite the training loop yourself - I hope the paper mentions all the important stuff and - without considering speed - could be done in kinda short time.

What could be the reason for missed detections in the below images?

I see two main reasons. Line size, see my answer to your 1st question, and domain, since the net was trained on natural images (see Wireframe dataset samples), but your domain is far from it.