evaluation.md

Evaluation

We provide scripts to run the training jobs in benchmark/synthetic/scripts. These can be launched as below.

Prediction Frame Alignment

Since the predictions using multi-view without pose supervision are arbitrary aligned, we first need to compute the optimal translation and rotation to bring them to the ShapeNet frame.

cd benchmark
cat synthetic/scripts/compute_alignment | bash

Rotation Evaluation

cd benchmark
cat synthetic/scripts/eval_rot | bash

The median errors and prediction accuracy for various settings will be saved in cachedir/resultsDir/pose/shapenet.

Shape Evaluation

To compute the mean IoU across categories for the different experiments -

cd benchmark
#save shape predictions
cat synthetic/scripts/eval_shape | bash

#evaluate shape predictions
cd benchmark/synthetic/
matlab -nodesktop -nosplash
>> evalShape