The new model doesn't work well.

[Wenzhengchina]

Hello,

Thanks for your interesting work. I tried to implement the pre-trained model on my own data. And I found some problems:

Firstly, I want to confirm the difference in the training process between the old and new model; you only fixed the bug of position encoding (temp_bug_fix=True) when you trained the new model (indoor_ds_new.ckpt), am I right?

Then, I did the evaluation of pose estimation on my dataset using your models. I found that the old model (outdoor_ds.ckpt) with temp_bug_fix=False works much better than the new model (indoor_ds_new.ckpt) with temp_bug_fix=True whether indoor or outdoor scene.

So, I'm curious as to why this is happening. Is there any other different setting in the training process? Generally, I think the performance shouldn't change a lot if you only revise the position encoding.

Can you explain the above problems? Looking forward to your reply.

Best,

[ghost]

I think the difference is caused by training dataset. As for my experience, outdoor model generally performs better than indoor model despite the positional encoding bug.

[oxyhexagen]

Then, I did the evaluation of pose estimation on my dataset using your models. I found that the old model (outdoor_ds.ckpt) with temp_bug_fix=False works much better than the new model (indoor_ds_new.ckpt) with temp_bug_fix=True whether indoor or outdoor scene.

Hi, may i ask how did you evaluate inferences on your own datasets? By using related camera pose decomposed from kpts as criterion？or comparing with corresponding points of annotation ground truth? I noticed that the new indoor ds weight matched much sparser kpts than the old outdoor ds weight, but i cannot assert which one is more precise.