cogtoolslab / physics-benchmarking-neurips2021

Repo for "Physion: Evaluating Physical Prediction from Vision in Humans and Machines", presented at NeurIPS 2021 (Datasets & Benchmarks track)
MIT License
59 stars 3 forks source link

generate & validate model test data #23

Closed judithfan closed 3 years ago

judithfan commented 3 years ago

To test the models on the same physical scenarios that are presented to humans, we need to re-generate exactly those 150 x 8 HDF5s, except where the target objects are painted in red/yellow.

Example to human input data for dominoes (but not red/yellow painted): https://github.com/cogtoolslab/human-physics-benchmarking/blob/master/experiments/dominoes_pilot/human-physics-benchmarking-dominoes-pilot_production_1_experimental_stims.json Subdirectory containing code to run all human experiments, including human-facing metadata: https://github.com/cogtoolslab/human-physics-benchmarking/tree/master/experiments