How do I choose the forward (x+) direction?

[xyysname]

Hello, I read the manual and learned the steps to generate the forward (x+) direction. But when I operate, I find that I can generate many values of the forward (x+) direction, right? So what I want to confirm now is whether the value of the forward (x+) direction can be arbitrarily determined? Or is it unique?

[liuyuan-pal]

Hi, you may feel free to define an x+ direction, which does not affect the performance too much. However, x+ affects how the 3d bounding box looks like in the finale results.

[xyysname]

hello，Now I get the pose of my target object through the experiment step of mouse.mp4. How should I evaluate whether the predicted pose is correct?

---Original--- From: "Yuan @.> Date: Thu, Dec 22, 2022 22:11 PM To: @.>; Cc: @.**@.>; Subject: Re: [liuyuan-pal/Gen6D] How do I choose the forward (x+) direction?(Issue #42)

Hi, you may feel free to define an x+ direction, which does not affect the performance too much. However, x+ affects how the 3d bounding box looks like in the finale results.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

[liuyuan-pal]

The evaluation requires manually labeling the pose. You may refer to https://github.com/liuyuan-pal/Gen6D/issues/38#issuecomment-1333131392

[xyysname]

Hello, I want to apply this algorithm to the pose estimation of a specific industrial object.

I conducted my own experiment: I took a video of the object, and then detected it according to the steps of mouse experiment, but found that its pose detection was not stable at some angles, and even the detection frame in the picture had a large deviation.

What are the possible reasons for such a large deviation? Is the angle of the object in my ref video insufficient? How can I reduce such large deviations?

This object is not common in reality, so I need to retrain the network so that the network can better predict its arbitrary position and pose? If I retrain the network, can it detect the pose of this object stably?

[liuyuan-pal]

Hi, you may post your intermediate results here (refer to https://github.com/liuyuan-pal/Gen6D/issues/4#issuecomment-1186136894) so I can figure out which part is not working well on your object. Retraining on your object is helpful, you may also check other object pose estimator like PVNet for instance-specific pose estimation.

[xyysname]

Hello, I would like to ask how to make my own detection target dataset? I plan to make my test object into the format in the linemod dataset, and then I find that your linemod dataset has many more files. Are these files necessary for training? If not all of them are necessary, can you tell me which ones are necessary, and how should I generate them? I would be very grateful if I could receive your reply.

[liuyuan-pal]

Hi, you may notice that the necessary thing is to implement the interfaces of the BaseDatabase. Note we use two different databases, one is for the reference sequences while the other is for the test. Reference one is required to have known poses.