Hi, thanks for this very nice work. I read the paper, but I have some questions:

1. Is your "LineTR_weight.pth" trained on scannet like superglue? As far as I know scannet does not have the ground truth of the line.
2. I noticed your fig2, just one figure input the transformer and line signature networks. Does it mean that training transformer and line line signature networks does not require the ground truth of scannet, just input line and point that have been detected?

Thank you for your work. When i view your code,I noticed that there is no code for homography. Cloud you teach me how to achieve the result like Figure 5 in your paper？I have sent you an email. Looking forward to your reply.

Hi @yosungho,

Thanks for this very nice work and for sharing the code, it is very insightful!

When studying the code, I noticed something weird that might be a bug. It is located in the KeylineEncoder at line: https://github.com/yosungho/LineTR/blob/afd89effcc081fe75105ea9c3598fdc5e91b6179/models/line_transformer.py#L121 It seems that the line descriptor is not updated at each iteration, since the output of the desc_layer has a different name than the input (desc vs enc_output). This does not trigger any error from the compiler, but the effect is that only one iteration of desc_layers is effectively used... The previous iterations are just overwritten each time.

Let me know if I got a wrong understanding of this line. Otherwise I think that the correction would be the following:

for desc_layer in self.desc_layers:
   desc, enc_slf_attn = desc_layer(desc, slf_attn_mask=mask)
   enc_slf_attn_list += [enc_slf_attn] if return_attns else []
enc_output = desc

Hi! I'm very interested in your work because recently I want to use line descriptors for homography estimation. I noticed that you implemented the code of the paper "Combining line and point correspondences for homography estimation". But I did not find out the code of that part in this repository for homography estimation. Could you provide the code for that part?

Hi, thanks for your great work. I wonder that how to test the model trained on my own dataset. I find that the match_line_pairs.py load superpoint model and linetr model but not my model_best.pth.tar? So how could I use it to infer.

Hello, when i run the original code, it always occurs the ZeroDivisionError, it also happens after i change the dataset. The reason is that the TP and FP are both 0 at the same time, so the precision is very low. It seems like the network is not converging. What should i do? Traceback (most recent call last): File "train.py", line 263, in <module> main() File "train.py", line 93, in main is_best = run_epochs("val", conf, val_loader, model, None, logger, epoch, device, lr_scheduler) File "train.py", line 233, in run_epochs result.evaluate(pred, batch_data, loss.item(), batch_index) File "/home/linux/catkin_ws/LineTR-train/evaluations/metric.py", line 40, in evaluate precision, recall, f1_score = self.eval_pr.get_precision_recall(score_pred, score_gt) File "/home/linux/catkin_ws/LineTR-train/evaluations/evaluate_pr.py", line 31, in get_precision_recall precision = np.average(np.nan_to_num(TP / (TP+FP)))*100. ZeroDivisionError: division by zero

Hello @yosungho, thank you for your paper!

Would it be possible for you to, please provide the code for the generating ground-truth line correspondences and 3D linemaps part on InLoc, so it would be possible to test visual localization results? I have sent you an email (address provided in the paper) with the same question.