3DHP Generalization Evaluation

[GR1778]

Hello. I would like to ask whether the generalization evaluation for your 3DHP dataset follows the testing protocols established by previous open-source works. Also, why did you not adopt the mainstream approach of evaluating on the training set using models trained from scratch?

[xiu-cs]

Hi @GR1778, thanks for your interest in our work! Let me address your two questions.

For the evaluation on the 3DHP dataset, I have not seen any other monocular single-frame-based repositories release their 3DHP
evaluation code; most only share the Human3.6M processing pipeline. For the corresponding code I provided, I built it mainly based on 3 sources: (a) the official 3DHP test set, which ships its evaluation metrics as MATLAB scripts (test_util/); (b) the P-STMO codebase; and (c) private processing code kindly shared by a senior researcher in the field.

I have checked our implementation against the official 3DHP MATLAB evaluation, and it follows the same protocol and metric definitions. I hope releasing this code helps make 3DHP evaluation more transparent and consistent across projects, rather than each work relying on its own private pipeline.

For your second question, I think your question is why we did not train the model on the 3DHP training set. Video-based methods usually train and test models on 3DHP. For single-frame methods, researchers usually evaluate cross-dataset generalization by directly testing a model pre-trained on Human3.6M.

Hope this helps clarify your questions.