Hello, thank you very much for open‑sourcing the dataset and complete code.
When reproducing and retraining NeuralILT, I encountered a boundary artifact: obvious white lines appear around the edges of mask images output by model inference. I only changed the training scheme by adopting early‑stopping. This problem does not occur when testing directly with the official weights provided in the project.
I would like to ask whether these boundary white lines stem from convolution padding, training strategy, or other boundary‑processing factors. Were more pre‑training epochs used during model training, so that the peripheral regions of the mask can learn from ground‑truth masks and avoid generating spurious patterns?
I would greatly appreciate it if you could offer some troubleshooting suggestions. Thank you!
Hello, thank you very much for open‑sourcing the dataset and complete code.
When reproducing and retraining NeuralILT, I encountered a boundary artifact: obvious white lines appear around the edges of mask images output by model inference. I only changed the training scheme by adopting early‑stopping. This problem does not occur when testing directly with the official weights provided in the project.
I would like to ask whether these boundary white lines stem from convolution padding, training strategy, or other boundary‑processing factors. Were more pre‑training epochs used during model training, so that the peripheral regions of the mask can learn from ground‑truth masks and avoid generating spurious patterns?
I would greatly appreciate it if you could offer some troubleshooting suggestions. Thank you!