Own dataset training results are not accurate

[yuanyesjtu]

Thanks to the authors for the excellent work. We have successfully reproduced the experiments above on datasets LineMod and YCB. However, when we tried to apply the model to our own dataset, the training results were not accurate. Our dataset contains 1 model, 1000 scenes. The depth map, color map, and mask map are as follows: This is model (We have compared our model with the model of the LM dataset, and their scales are similar. The selected keypoints are indeed on the model.):

This is scene (The scale is 640*480.): This is the training process:

This is our training results:

We compared our own data with LM data, and the similarities and differences are as follows:

1. There is little difference between the models;
2. The dpt_mm of the LM data contains 270,000 non-zero points, and our dpt_mm contains 80,000 non-zero points.
3. LM's kp_targ_ofst and ctr_targ_ofst contain about 100 points, and our kp_targ_ofst and ctr_targ_ofst contain about 600 points.
4. Our nrm_map is significantly sparser than that of the LM dataset, such as: This is our nrm_map: This is LM nrm_map:

Can you point out the reason for our error? Thank you very much.

[bjutlss]

Hello, have you solved your problem? May I ask you some questions? @yuanyesjtu