NVlabs / Deep_Object_Pose

Deep Object Pose Estimation (DOPE) – ROS inference (CoRL 2018)
Other
1.03k stars 287 forks source link

How to test model accuracy #386

Closed brumocas closed 1 month ago

brumocas commented 1 month ago

Hello again, I have been training DOPE on custom objects and I am getting some results. The model is the following: Screenshot from 2024-09-26 16-37-57

The inference results are like the following images: 00123 00135 00135 00135 00190 00205 00299 00331 00379

I want to implement this in a real robot application but first i need to evaluate model accuracy How can I generate a dataset like the one you did for the cracker box, with ground truths? Thanks in advance :) image image image (...)

TontonTremblay commented 1 month ago

I wrote a script that in nvisii renders the data on a table like this. https://github.com/TontonTremblay/nvisii_mvs?tab=readme-ov-file#falling-object-scene I think this is what I used, there might be some options to make sure the object is upright.

brumocas commented 1 month ago

It worked for now :muscle: Really appreciate your help

TontonTremblay commented 1 month ago

Cool show me some renders you did :P would love to see more, I like when people are using my broken code I wrote :P

brumocas commented 1 month ago

Sure :)
I managed to get the table and the object in the scene after playing around with the config. The hdri background scene is not showing up for some reason, any idea how to solve this? Is the code only using the light from the hdri file? Can i change the object position in the scene, for example upside down?

00000 00030 00058

brumocas commented 1 month ago

Solved