Post-Processing¶

Storage¶

Ensure that data is backed up to prevent data loss during post-processing.
- Protect against accidental writes / deletions:
  - Linux: chmod -R a-w
- Store backup in a safe place. I.e. do not travel with backup disks.
Control access to data
- Protect against unauthorized access:
  - Linux: chmod -R o-rwx
Do not operate with the raw (or even worse backup) data. Separate corpus from derived data/features.
- Also helps to restart automated processing from scratch in case of errors because directories are separated.

Always: Automate as much as possible.

Extract trials which fulfill quality criteria:
- Recording complete?
- Use validation script
- Ignore trials/recordings where issues occurred during the recording session (as indicated in the recording log).
Extract regions of interest from videos etc.
Conversion to target format:
- Decide on target formats (video, audio, CSV? etc) depending on usage patterns (h264+aac is windows, mac and linux compatible).
- Generate multiple resolutions/sizes in case it helps quick annotation/processing

Think about the required accuracy of synchronization.
- Think about how to validate the accuracy.

Define and document an annotation scheme based on literature and with relation to hypotheses.
Think about how to proof annotation reliability (interrator agreement etc.)
- Use multiple persons to annotate the data in parallel and isolation.
Train annotators.

Have a look at the Dataset Processing Project for some useful scripts.

Automate audio and video analyses (using existing tools)
- Blackframe detection to synchronize videos containing black frames (e.g. ffmpeg)
- Detection of known sound patterns (e.g. clapperboard, robot utterance, via praat and cross-correlation to reference audio signal)
- Estimation of temporal offsets between cameras (cross-correlation)
- Clapperboard detection using Vicon markers (distance of two markers)
Align videos to system logs by e.g. recording one audio channel via middleware or more sophisticated solutions

Tools: