Model Directory Structure
/models/model_dir/
├── config.yaml
|
# Singleview
├── predictions.csv
# Multiview
├── predictions_view0.csv
├── predictions_view1.csv
|
├── tb_logs/ ...
├── video_preds/
│ ├── session0_view0.mp4/
│ | └── predictions.csv
│ └── session0_view1.mp4/
└── image_preds/
└── CollectedData.csv/
└── predictions.csv
Detailed descriptions
tb_logs/: model weightsvideo_preds/: predictions and metrics from videos. The config fieldeval.test_videos_directorypoints to a directory of videos; ifeval.predict_vids_after_trainingis set totrue, all videos in the indicated direcotry will be run through the model upon training completion and results stored here.video_preds/labeled_videos/: labeled mp4s. The config fieldeval.test_videos_directorypoints to a directory of videos; ifeval.save_vids_after_trainingis set totrue, all videos in the indicated direcotry will be run through the model upon training completion and results stored here.predictions.csv: predictions on labeled data. The right-most column records the train/val/test split that each example belongs to.predictions_pixel_error.csv: Euclidean distance between the predictions inpredictions.csvand the labeled keypoints (in<YOUR_LABELED_FRAMES>.csv) per keypoint and frame.
We also compute all unsupervised losses, where applicable, and store them (per keypoint and frame) in the following csvs:
predictions_pca_multiview_error.csv: pca multiview reprojection error between predictions and labeled keypointspredictions_pca_singleview_error.csv: pca singleview reprojection error between predictions and labeled keypoints