SizheAn
frames are different in radar feature and 3D pose label
I try to use subject{i}_featuremap.npy as training data and 3D pose label as label to training , just like MARS dataset . But I found the frames in each subject{i}_featuremap.npy are different with 3D pose label in subject{i}_all_labels.cpl['naive_gt_kps']
(6384, 14, 14, 5) (6529, 3, 17) (7150, 14, 14, 5) (7205, 3, 17) (6982, 14, 14, 5) (7042, 3, 17) (6893, 14, 14, 5) (6956, 3, 17) (7219, 14, 14, 5) (7243, 3, 17) (6906, 14, 14, 5) (7014, 3, 17) (6743, 14, 14, 5) (8944, 3, 17) (6817, 14, 14, 5) (6876, 3, 17) (6704, 14, 14, 5) (6817, 3, 17) (6911, 14, 14, 5) (6953, 3, 17) (7101, 14, 14, 5) (7334, 3, 17) (6526, 14, 14, 5) (6761, 3, 17) (6876, 14, 14, 5) (7026, 3, 17) (6812, 14, 14, 5) (6981, 3, 17) (6557, 14, 14, 5) (6787, 3, 17) (6722, 14, 14, 5) (6924, 3, 17) (6387, 14, 14, 5) (6602, 3, 17)
I'm not sure how to get aligned dataset with labels
Can you check the keys: 'radar_avail_frames'? It specifies the available frames of radar modality. Then you can choose the available ones from the groundtruth.
I also had the same issue as the person above. The radar_avail_frames have keys [132, 6808] so total available frames are 6940 in this modality. Also, it is not the 'naive_gt_kps' rather it should be the 'gt_avail_frames' as that contains available frames for 3d human joints ground truth. Is this what you're refering to when you're saying to choose the available ones from groundtruth?
Sorry for the confusion and I will update a better readme later. Here I'm showing an example:
For this subject, there are 6529 frames in total. so 6529 for the shape of gt. The available radar frame starts from index 98 and ends at 6481. As a result, total frames of available radar is 6481-98+1=6384. And this is identical to the shape of radar features. I hope this resolves your problems.
Thank you so much for the clarifying this issue. This makes more sense. Thanks