search

TRI-ML / vidar

Other
554 stars 65 forks source link

Incorrect predicted depth of closest in path objects #33

Closed akashchavan15 closed 1 year ago

akashchavan15 commented 1 year ago

I have used monocular depth estimation for my own data. I see that the depths for closest-in-path objects are not correct. Can someone help me to fix this?

Please find the attached images for more details.

Incorrect_depth Incorrect_depth_1

akashchavan15 commented 1 year ago

can anyone please provide some suggestions?

akashchavan15 commented 1 year ago

@RaresAmbrus Could you please provide some suggestions?

ivasiljevic commented 1 year ago

The photometric loss makes a static scene assumption; if we train on enough data in practice we get models that can predict reasonable depth on dynamic objects (e.g. by generalizing from parked cars to cars in front of the ego-vehicle). If your dataset is on the smaller side, you're going to get missing depth like that because the dynamic object is masked out.

akashchavan15 commented 1 year ago

@ivasiljevic Thank you for your reply. I have 60000 images in my training data. What is the best practice I can follow to filter out dynamic scenes from my dataset? I removed static frames.