Update FEEL documentation

README.md

@@ -18,6 +18,28 @@ repository is archive-oriented.
 `Kitchen/participant_mapping.tsv` maps anonymized Kitchen IDs back to the
 original local recording names.
 
+## Getting started
+
+The full dataset is approximately 600 GB. We recommend starting with
+`Kitchen/P12` (the Yianni recording), which is one of the higher-quality
+recordings, before downloading the remainder of the dataset.
+
+## Force preprocessing and sensor quality
+
+The piezoresistive sensors are noisy. [`preprocess.py`](preprocess.py) contains
+the preprocessing pipeline used to derive contact ground truth from the raw
+sensor measurements. This pipeline is intended for contact estimation; work
+that directly predicts force magnitudes will require a somewhat different
+preprocessing pipeline.
+
+Some fingers experienced sensor drift or sensor malfunction during a
+recording. A combination of manual and automatic review was used to filter out
+the affected measurements. In the processed CSV files, the corresponding force
+columns are set to `NaN` for the affected intervals. Consequently, some fingers
+are absent from the supplied visualizations. Learning pipelines should treat
+these values as missing data rather than as zero-force observations, and should
+mask or otherwise handle them during training and evaluation.
+
 ## Known data issue
 
 Lab participant `P21` contains 4,707 source JPEG files that OpenCV cannot