search

TUI-NICR / ESANet

ESANet: Efficient RGB-D Semantic Segmentation for Indoor Scene Analysis
Other
239 stars 52 forks source link

I have a question about your ESANet model inference time. #35

Open suesuekkim opened 2 years ago

suesuekkim commented 2 years ago

Firstly, I really appreciate about your clean and kind code.

I have trained ESANet model with SUNRGBD dataset and scenenet dataset. For the SUNRGBD dataset, when i implemented the code prepare_dataset.py in the src colder, the refined depth data was generated, and when i train the depth mode to 'refined', it was trained with the refined depth dataset. Also when i downloaded scenenet dataset, the depth file was refined depth. And I have a question here, in the paper of this model, is the inference time that you mentioned in Fig 4. and Fig 5. include the data pre-processing time (making the raw depth to refined depth)? Without refining the depth data, is it able to implement in the real-time application?

Thank you for reading my issue.

mona0809 commented 2 years ago

Yes, we did train on refined depth data, as it is common practice. Still, you can use the model and apply it on "raw depth" samples. However, you can also train on raw depth data, which results in only slightly lower mIoU. For application, I think, you will not notice the difference.