About dataset preprocessing: how to match the original videos with annots.npz in mpi_inf_3dhp

[tinatiansjz]

ROMP is an impressive work indeed! Thanks for sharing the code. Besides, it's admirable of you to deal with all the issues in time.
I have a little question about preprocessing of MPI_INF_3DHP dataset.
I examined the img_names field in dataset/mpi-inf-3dhp/cluster_results_noumap_mpiinf_kmeans.npz and found out that there exists video_[9,10] in every S[1-8]_Seq[1-2]. It confuses me as the original MPI_INF_3DHP only contains video_[1-8] in S[1-8]_Seq[1-2], as illustrated below.

I wondered how to match the original videos with annots.npz or cluster_results_noumap_mpiinf_kmeans.npz that is officially provided. To put it another way, to which original videos that the frames with filenames containing video_[9,10] correspond?
It'd be appreciated if you could give me some advice. Thank you for your time.

[Arthur151]

Thanks for your kind words!
The code about extracting frames from video is placed at here.
About the MPI_INF_3DHP datasets, you may miss some videos. It contains 14 videos (0-13) in each Seq. Please refer to this introduction. I drop the unusual view of 11-13 and leave 0-10 for training.

[Arthur151]

No feedback

[tinatiansjz]

Thank you for your notes and I apologize for the late reply.
I looked at the officially provided MPI-INF-3DHP download link on the project page and re-checked the config.ig.
It turns out that download_extra_wall_cameras and download_extra_ceiling_cameras (Line 13-14) need to be set to 1. Then we can get the additional video_[9-13] in every S[1-8]_Seq[1-2].

Thanks again for your time. The question has been solved!

[Arthur151]

Thanks for sharing this with other developers.

[liangwx]

请问cluster_results_noumap_h36m_kmeans.npz和cluster_results_noumap_mpiinf_kmeans.npz是如何生成的，没有找到生成代码

[Arthur151]

这个涉及到我们另一篇论文使用的姿态聚类方法（arxiv论文的Fig.12），相关代码会在论文接收后开源。

[liangwx]

请问还有annots_train2014.npz和annots_val2014.npz的生成过程和代码呢？

[Arthur151]

在这里。每个数据集的annots.npz的生成函数就是各自的pack_data