wf-process-pose-data 6.3.0

Tools for fetching, processing, visualizing, and analyzing Wildflower human pose data

# process_pose_data

Tools for fetching, processing, visualizing, and analyzing Wildflower human pose data

## Task list

• Switch parallel overlay code back to imap_unordered() (for less chunky progress bars) but sort output before concatenating

• Ensure that all visual specs (colors, line widths, etc.) propagate to video overlay

• Add drawing primitive to wf-cv-utils for text with background

• Use new text-with-background drawing primitive for pose labels

• Add timestamp to video overlays

• Consider getting rid of geom_render module

• Fix up progress bars so we can have task-level and segment-level progress bars without logging in notebook

• Fix up progress bars so they work properly outside of notebook with multiprocessing

• Add indentation/labeling on segment-level progress bars

• Clean up argument ordering in reconstruct_poses_3d_alphapose_local_time_segment

• Add poses_2d_json_format option to reconstruct_poses_3d entry point

• Add endpoints for all processing functions

• Add ability to write locally generated object IDs to Honeycomb

• Create separate workers for uploading to Honeycomb 3D poses, 3D pose tracks, interpolated 3D pose tracks, 3D pose identification, 3D pose track identification

• Dockerize pipeline

• Set up pipeline for Airflow

• Add additional machinery for checking UWB data integrity (e.g., duplicates)

• Retool generate_inference_metadata_reconstruct_3d_poses_alphapose_local to exclude cameras without calibration data

• Make function to delete Honeycomb inference executions

• Make function to delete local inference metadata

• Make function to delete local 3D pose files

• Make function for deleting local 3D pose data

• Rewrite all log messages so formatting isn’t called if log isn’t printed

• Make functions handle empty poses (all keypoints NaN) more gracefully (e.g., score_pose_pairs(), draw_pose_2d())

• Make visualization functions handle missing fields (e.g., pose_quality) more gracefully

• Figure out inconsistent behavior of groupby(…).apply(…) (under what conditions does it add grouping variables to index?)

• For functions that act on dataframes, make it optional to check dataframe structure (e.g., only one timestamp and camera pair)

• For functions than iterate over previous functions, making naming and approach consistent (e.g., always use apply?)

• Add keypoint_categories info to pose models in Honeycomb?

• Be consistent about whether to convert track labels to integers (where possible)

• Remove dependence on OpenCV by adding necessary functionality to cv_utils

• Consider refactoring split between video_io and cv_utils

• Fix up cv_utils Matplotlib drawing functions so that they accept an axis (or figure, as appropriate)

• Fix up handling of background image alpha (shouldn’t assume white background)

• Fix up _y_ axis inversion for images (go back to cv_utils?)

• Add option of specifying Honeycomb client info for visualization functions that require Honeycomb

• Reinstate sns.set() for Seaborn plots without making it spill over into non-Seaborn plots (see [here](https://stackoverflow.com/questions/26899310/python-seaborn-to-reset-back-to-the-matplotlib))

• Refactor code in visualize to make it less repetitive (same pattern over and over for [verb]_by_camera)

• Fix up legend on pose track timelines

• Add visualization for number of poses per camera per timestamp

• Replace cv.triangulatePoints() to increase speed (and hopefully accuracy)

• Get pose video overlays working again (for data with track labels)

• Rewrite geom rendering functions to handle the possibility of no track labels

• Rewrite function which overlays geoms on videos so that user can specify a time span that it is a subset of the geoms and/or the video

• Make all time inputs more permissive (in terms of type/format) and make all time outputs more consistent

• Be consistent about accepting timestamp arguments in any format parseable by pd.to_datetime()