search

Lavreniuk / EVP

Metric depth estimation from a single image and referring segmentation
https://lavreniuk.github.io/EVP/
MIT License
59 stars 5 forks source link

Visualize KITTI result and KITTI_best_model.pth #15

Open RuiTianHIT opened 1 month ago

RuiTianHIT commented 1 month ago

@Lavreniuk Hello author, This is a high-quality work and we are very interested in this work The reproduced results are consistent with those in the paper based on the KITTI dataset. However, we encountered difficulties during visualization and the visualization results were not ideal. (1) We showed our visualization results. We don’t understand why the evaluation indicators are excellent, but the visualization results are not ideal. 0000000006

(2) I hope the author can provide best_model.pth on the KITTI dataset so that we can better understand whether it is a problem of visualization or model parameters (the model we trained ourselves is in the KIITI dataset)

Looking forward to your answer, thank you very much for your help! ! !

Lavreniuk commented 1 month ago

@RuiTianHIT, I am delighted to hear that everything works and the scores you reproduced are high and quite close to those written in the paper. I have uploaded the best KITTI model for you here: https://huggingface.co/spaces/MykolaL/evp/tree/main/depth

I have checked it on visualization, and it works well. Please make sure that you have good visualization for this model, and then switch to your weights. data_0000000005 data_0000000007

RuiTianHIT commented 1 month ago

@Lavreniuk With your help, this issue has been resolved and the visualization effect is of high quality. Thank you very much. I have a question to discuss whether there is a situation in depth estimation tasks where the evaluation metrics are good (d1 or RMSE), but the visualization effect is not good. I think it may exist because the nussence dataset is also a classic depth estimation benchmark, GT only exists in some areas, so does the model learn local solutions?

RuiTianHIT commented 1 month ago

@Lavreniuk Could you provide the loss for each epoch when training with KITTIdataset? And the results output for each epoch? Recently, we have encountered problems during training and would like to compare and further identify the cause of the error.