scores in resulst and bop19_average_recall

[cats0212]

Question 1:I used ycbv-jwpvdij1.compact.ckpt(a trained model that you provided) to infer test datasets in ycbv(python -m surfemb.scripts.infer), then python -m surfemb.scripts.misc.format_results_for_eval, the score in results all is negative, for example，-0.339 , -0.401.Is that normal？ A:scene_id B:img_id C:est_obj_id D: score.

[cats0212]

Question 2:Through your trained model,I get the file ycbv-jwpvdij1-refine-pose-score_ycbv-test.csv. Then python ./bop_toolkit/bop_toolkit_lib/eval_bop19.py get a result { "bop19_average_recall": 0.5618441264451451, "bop19_average_recall_mspd": 0.7611933058452582, "bop19_average_recall_mssd": 0.4954887218045113, "bop19_average_recall_vsd": 0.42885035168566576, "bop19_average_time_per_image": 4.59347986459732 } But average_recall(ycbv) is 0.653 in your paper.Is there something wrong with me?thank you

[cats0212]

Question 3: ycbv-jwpvdij1.compact.ckpt,the .ckpt is trained purely on synthetic data？or on real data for ycbv.

[rasmushaugaard]

1) The scores are log-likelihoods, so they're supposed to be negative.

2/3) The released models are those used for the BOP evaluation, only trained on the synthetic PBR images. I just re-ran inference with the ycbv-jwpvdij1.compact.ckpt model and got 0.645, similar to the 0.647 published on BOP.

Based on the logs from the other issues, you've opened, it looks like you're using python3.7, which in itself shouldn't be an issue, but the conda environment I have published is with python3.8. Have you tried that environment?

[cats0212]

ok,thank you,I will try it.