https://lear.inrialpes.fr/local/wiki/ 2023-09-06T10:25:59+02:00 https://lear.inrialpes.fr/local/wiki/ https://lear.inrialpes.fr/local/wiki/lib/tpl/dokuwiki/images/favicon.ico text/html 2017-01-19T18:03:35+02:00 https://lear.inrialpes.fr/local/wiki/doku.php?id=nips_3d_pose:main&rev=1484845415&do=diff NIPS 3D pose estimation Work-in-progress This documents how to use a slight rewrite of Gregory's NIPS 3D pose estimation code. * Train a model * Test an existing model * Plot error curves * Data format (.h5) List of relevant files: * source_common_deps.sh * test_caffe_model.sh * test_run_once_caffe.sh text/html 2016-12-19T13:39:49+02:00 https://lear.inrialpes.fr/local/wiki/doku.php?id=nips_3d_pose:plot_error_curves&rev=1482151189&do=diff [[nips_3d_pose:main|nips_3d_pose:main]] ====== Plot error curves ====== ===== plot_error_curves.sh ===== This script runs on multiple iterations of a trained model (eg. from 2000 to 50 000 iterations, stepsize=2000). X_axis = #iterations. The following curves are plotted: * classification error [0; 1] * absolute error (mm) * transformed error (mm) * 2D absolute error (px) text/html 2016-12-19T13:34:24+02:00 https://lear.inrialpes.fr/local/wiki/doku.php?id=nips_3d_pose:test_model&rev=1482150864&do=diff [[nips_3d_pose:main|NIPS 3D pose estimation]] ====== How to test an existing model ====== ===== test_caffe_model.sh ===== This script runs a trained model on a given test-set, and outputs absolute error (mm), transformed (mm) and pixel error (on 220x220 images). inputs: - caffemodel trained - clustering used during training (matlab file) - network definition file (DEPLOY version of trained network) - test set (.h5 file, must contain '/data' and '/joints3D_centered') depe… text/html 2017-01-19T18:05:56+02:00 https://lear.inrialpes.fr/local/wiki/doku.php?id=nips_3d_pose:train_model&rev=1484845556&do=diff [[nips_3d_pose:main|NIPS 3D pose estimation]] ====== How to train a model ====== ===== train_caffe_model.sh ===== This script trains a model from a given solver file. inputs: - solver file (net architecture, list of training samples) - (optional) reference model (for fine-tuning etc) dependencies/called functions: - source_common_deps.sh : Caffe path, Matlab path, etc.. - caffe_train_from_reference.py : Caffe wrapper when a reference model is used