search

NVlabs / Deep_Object_Pose

Deep Object Pose Estimation (DOPE) – ROS inference (CoRL 2018)
Other
1.01k stars 283 forks source link

How to identify the hollow model better #108

Open xdw-123 opened 4 years ago

xdw-123 commented 4 years ago

I've been able to identify my own training model. But the model is a kind of hollow structure, which requires too much light for the scene in the process of recognition. How to identify any environment? The measures include the improvement and requirements of changing the relevant parameters of training code or synthesizing data.

TontonTremblay commented 4 years ago

Do you mind sharing some examples, like images?

xdw-123 commented 4 years ago

Hello, I have uploaded a group of synthesized data and physical pictures. I hope you have some suggestions. How to recognize any scene?URL：https://github.com/xdw-123/Deep_objcet_Pose_text-Image.git

TontonTremblay commented 4 years ago

I did some testing on recognizing chairs with similar structure, e.g., thin legs. The network had no problem detecting the chair. I would think DOPE should work fine with your data (training + detection). Is it not the case?

xdw-123 commented 4 years ago

I think you are right. When you run rviz for simulation, there is frame dropping. How to speed up the recognition speed? My computer configuration: rtx8000, i9 7900