Inference on MEVA Dataset

[mateuszk098]

Hi,
I'm trying to run an inference notebook on the MEVA dataset -> https://mevadata.org/. I'm using two indoor cameras, i.e. G299 and G330 including their calibration parameters -> G299-params and G330-params. I prepared JSON file like yours, but using these parameters. However, I'm not sure how to interpret the "k" and "p" parameters. In the MEVA dataset, in calibration files, there is only one 2-element vector: "D - Nd vector of lens distortion parameters (using OpenCV conventions)". Should I set this vector as "k" or "p", or maybe not at all? The second question is about SPACE_SIZE and SPACE_CENTER. As far as I understand the SPACE_SIZE denotes dimensions of a 3D bounding box which covers the entire motion space, and SPACE_CENTER is its centre. So for example in the case with MEVA, I set (12000, 12000, 4000) and (6000, 6000, 2000) correspondingly but this does not seem to work. Moreover, I don't understand why in the panoptic config file, the SPACE_CENTER has a negative value for one dimension? The last question is about rotation and translation matrices for calibrated cameras. As can I see, the values in the JSON file you provided for the Campus dataset are completely different than one can find when downloading this dataset. Only the last row for R is the same. My question is, what exactly is the transformation you have used and why? As an example, here are the results I got for selected MEVA images along with calibration parameters and input images.
calibration_school.json

[JJ-Luov]

hello, did you figure out these questions?