accuracy issue while detecting vertical lines using pre-trained model

[trohit920]

Hello,
I have tried the DHT pre-trained models to detect vertical catenary lines but the accuracy is not great. Please check the attached image files. How do you think i should solve this issue, shall i retrain the model on my custom data?
Any suggestion will be of great help.

[HassanBinHaroon]

Hello, I have tried the DHT pre-trained models to detect vertical catenary lines but the accuracy is not great. Please check the attached image files. How do you think i should solve this issue, shall i retrain the model on my custom data? Any suggestion will be of great help.

Hi @trohit920!

Your question is quite legitimate and requires in-depth knowledge plus experience with particular architecture.

First of all, I want to refer you to the actual authors of this implementation. Kindly visit the authors' repository here.

Second of all, what I remember about deep hough transform is, there should be a suitable threshold value that can be figured out practically for your own problem. The algorithm is working pretty fine in the above (shown) image, and its threshold (by default) is selected for high specificity instead of high sensitivity.

[HassanBinHaroon]

Hello, I have tried the DHT pre-trained models to detect vertical catenary lines but the accuracy is not great. Please check the attached image files. How do you think i should solve this issue, shall i retrain the model on my custom data? Any suggestion will be of great help.

Moreover, GENERALLY SPEAKING (not tested), training on the specific dataset yields better results on that particular dataset most of the time. So, I suggest you give it a try.