Table 1.

Summary of related work reporting volume estimation results on a per-drink basis.

Paper Identifier	Sensing Modality	Estimation Quantity/Approach	# of Subjects/ # of Drinks	User-Specific vs. Out-of-Subject Model	Best Reported Per-Drink Performance Metric
Amftet et al. [7]	Wearable magnetic coupling sensors on wrist and shoulder	Fill level classification (3 levels: full, medium, near empty)	3 subjects/ 810 drinks	User-Specific	72% classification accuracy
Mirtchouket et al. [18]	Earbud, two smart watches, smart glasses with embedded IMUs	Volume regression	6 subjects/ 285 drinks	Mixed (i.e., both user-specific and out-of-subject training data)	47.2% MAPE
Hamataniet et al. [14]: Lab-micro+ collection	Commercial smartwatch with embedded IMUs	Volume regression	16 subjects/ 1069 drinks	Out-of-subject (user-specific results reported for benchmarking)	58.9% MAPE
Hamataniet et al. [14]: Wild office dataset	Commercial smartwatch with embedded IMUs	Volume regression	16 subjects/ 178 drinks	Out-of-subject, with models trained on Lab-micro + data and ground-truth collected via commercial smart bottle	34.6% MAPE
Griffith et al. [19]:	Bottle-attachable IMU Sensor	Binary volume classification with median volume partition	64 subjects/ 1200 drinks	Mixed (i.e., both user-specific and out-of-subject training data)	29.2% classification error for median partition
Current Manuscript	Bottle-attachable IMU Sensor	Volume and fill level regression	84 subjects/ 1908 drinks	Out-of-subject	52.4% MAPE (volume regression)