search

DrSleep / DenseTorch

An easy-to-use wrapper for work with dense per-pixel tasks in PyTorch (including multi-task learning)
MIT License
62 stars 9 forks source link

Predicted depth maps aren't as expected. #10

Open onizuka94 opened 4 years ago

onizuka94 commented 4 years ago

I have trained the model using a custom dataset, where the GT_depth maps were generated using colmap. However after training, and running the inference notebook, this is what I get as output, the predicted depth maps seems more to be as normal maps. @DrSleep any idea or guidance on the matter ? plot_mod

DrSleep commented 4 years ago

hard to say without any colourbars. Does the depth metric become better as the training progresses?

onizuka94 commented 4 years ago

@DrSleep nope, The RMSE is stuck at 0.8xxx .

DrSleep commented 4 years ago

0.8xxx by itself does not tell me anything.

Since you are using a custom dataset, make sure that its structure is identical to NYUD since that the dataset which the training example is based upon. Concretely, make sure that you set depth_scale and ignore_depth to the correct values in your configuration. Defaults are here

13sunmin commented 3 years ago

How do you get the segmentation and depth of field results from the trained model?

13sunmin commented 3 years ago

I've trained the model. How can I get the inference notebook.Can you help me?