Do I need to provide initial camera calibration parameters?

[mosamdabhi]

Owing to the proposed "Pictorial structure based calibration", ideally I should just provide detected 2D as input and get the 3D structure if I am understanding the paper correctly.

Thus, if I get the 2D poses from df2d in the following structure (7 x 15 x 38 x 2), do I need to provide any other information (such as calib.pkl) in order for me to generate a 3D structure (15 x 38 x 3)? In other words, what information do I need to provide? Intrinsics, Extrinsics, or both? Or initial rough values for both?

Asking the above question in case I need to run on Human 3.6 M dataset, where I just have 2D detected poses from a 2D pose detector and nothing else.

[semihgunel]

Hi, The pictorial structure is used for refining the 3D pose, and not really for calibration. Yes, you need to provide initial calibration information (as you pointed out, you can provide it with calib.pkl file). You need both intrinsic and extrinsic matrices. However extrinsic can be rough and not exact.
For the h36m case, the real camera matrices are already given with the dataset, so you can just plug them. Let me know if you have more questions.

[mosamdabhi]

Hi @semihgunel,

It looks even after providing groundtruth camera matrices a[s well as clean groundtruth 2D ends up providing 3D structures that don't seem plausible. Am I missing something?