CMU-trained Weights and Cross dataset evaluation

[Samleo8]

In your paper under Experiments > CMU Panoptic Dataset, you noted that

We also conducted experiments to demonstrate that the learnt model indeed generalizes to new setups. For that we applied a CMU-trained model to Human3.6M validation scenes. (...) To provide a quantitative measure of the generalizing ability, we have measured the MPJPE for the set of joints which seem to have the most similar semantics (namely, ’elbows’, ’wrists’ and ’knees’). The measured MPJPE is 36 mm for learnable triangulation, and 34 mm for the volumetric approach, which seems reasonable when compared to the results of the methods trained on Human3.6M (16-18 mm, depending on the triangulation method)

I am trying to replicate some of the results, but was not exactly successful. I understand that the results are different when running H36M-trained on CMU validation (rather than the other way round as you did in the paper), but as can be seen from my test results in issue #76 , I get quite poor results. Granted, you did mention in your paper that the CMU-trained weights were the best because it could learn from truncated views, but it suggests that the H36-trained model is fairly poor when dealing with occlusions.

It would be great if you could provide weights for the CMU-trained model, as it would help validate the results in your paper, and also augment the strength/usability of your model. I believe that this model is will be great for auto-generating 3D ground-truth data for many of the community's datasets, but without the CMU-trained model, many in the community cannot use it practically for these purposes.

Also, how did you make a proper comparison between the joints when the CMU and H36M have vastly different joint positions? Is there code for this?

Thank you!