Thanks for your work. I tried training it on the monocular endoscopy dataset EndoSALM, but the results were poor.

[wza527]

The color images of the dataset are 320*320, as follows:

The real depth images of the dataset are saved in 32-bit .png format, as follows:

The data loader I wrote is as follows:





My evaluation results were very bad, and the results are as follows:

What do you think is the reason? I changed max_depth to 100

[noahzn]

Hi, I haven't worked on such dataset. What is the actual range of your ground-truth? Maybe 100 is too large for your dataset.

[wza527]

嗨，我还没有处理过这样的数据集。您的真实值的实际范围是多少？也许 100 对于您的数据集来说太大了。

Thanks for your reply. The actual range of my data set is 0-100mm. The unit of your code parameter is meter. How should I change it appropriately?
Thanks again for your quick reply.

[noahzn]

If the range is 0-100mm, please make sure that your evaluation code can correctly deal with this range.

[wza527]

1. Where do I need to change the evaluation code to ensure that my code can handle it correctly?
2. I ran test_simole.py and the results were not good. I used the pre-trained weights of ImageNet from your lite-mono-small, whose input size is 640192, and my data is 320320. Will this affect the training results?

[noahzn]

1. As you can see here the MIN and MAX are defined in the evaluation code. It can be a problem if your actual ground-truth is not in this range.
2. Any input size is compatible with the pretrained ImageNet weights.

[wza527]

1. 正如您在这里看到的，MIN 和 MAX 在评估代码中定义。如果实际的 ground-truth 不在此范围内，则可能会出现问题。
2. 任何输入大小都与预训练的 ImageNet 权重兼容。

I have changed it there.I think it's a problem with the training effect. But I don't know what causes the poor training effect.

[noahzn]

If your camera intrinsic is correct... How are the curves and visualizations during training? Have you checked them using tensorboard?

[wza527]

如果您的相机内在是正确的...训练期间的曲线和可视化效果如何？您是否使用 tensorboard 检查过它们？

The training curve changes are as follows:

The visualization effect is as follows:

[noahzn]

The visualization looks not bad. But your evaluation results are totally bad. I think the problem is still from your dataset. I saw several papers using Lite-Mono for endoscopy depth estimation and they reported good results.

[wza527]

可视化效果看起来还不错。但是你的评估结果完全糟糕。我认为问题仍然出在你的数据集上。我看到几篇使用 Lite-Mono 进行内窥镜深度估计的论文，它们报告了良好的结果。

OK. I am using the public EndoSLAM dataset, so it should be fine. Maybe there is something wrong with the loader I use to process the data. Can you tell me the title of the paper?

[noahzn]

https://scholar.google.com/scholar?hl=en&as_sdt=2005&sciodt=0%2C5&cites=6869562733583933958&scipsc=1&q=endoscopy&btnG=

[wza527]

Thanks for your help.

[noahzn]

You are welcome. Please update if you find the reason.

[wza527]

You are welcome. Please update if you find the reason.

OK