Skip to content

Latest commit

 

History

History
70 lines (48 loc) · 2.44 KB

README.md

File metadata and controls

70 lines (48 loc) · 2.44 KB
Raw

eval-mpii-pose

Scripts for evaluating results on the MPII human pose dataset. "Single person pose estimation" is the only type of evaluation currently supported.

Disclaimer: This is an unofficial repository, I am not from MPI and I was not involved in the creation of the dataset.

Input format

Predictions are expected to have the following format:

  • Must be a Matlab (.mat) or HDF5 (.h5) file
    • Must have one field, preds, which is the joint predictions tensor
    • Tensor size must be [2 x 16 x n] or [n x 16 x 2]
  • Must correspond to one of the following subsets: train, val, test
    • See annot/{train,valid,test}.h5 for which examples are in each of these subsets

Predictions produced by the following repositories meet these requirements:

Metrics

PCKh

The PCKh performance metric is the percentage of joints with predicted locations that are no further than half of the head segment length from the ground truth.

"PCKh total" excludes the pelvis and thorax joints from the calculation, presumably because they are very easy to predict given that the approximate person center is provided.

Scripts

evalMPII.m

Loads predictions from Matlab or HDF5 files and compares them with ground truth labels to calculate accuracy metrics (eg PCKh). You will want to edit getExpParamsNew.m to add new sets of predictions, and evalMPII.m to specify which predictions to include and which subset (train/val) to use.

prepareTestResults.m

Loads flat test set predictions and prepares them for submission.

Reference predictions

The preds/reference directory contains multiple validation set prediction files generated by established pose estimation models. You can compare against these predictions using evalMPII.m.

NOTE: Since the reference predictions are for the validation set, they are not compatible with the prepareTestResults.m script.

File origins

In order to keep evaluation in line with existing work, a lot of files in this repository were copied verbatim from other sources.