Source-linked AI summary
Doppio: A Dataset for Contactless Weight Estimation of Falling Particles
Simon Kiefhaber, Jan-Martin O. Steitz, Julia Grabinski, Christoph Reich, Paul Wagner, Max Zimmermann, Simone Schaub-Meyer, Stefan Roth
TL;DR
Powder-mass measurement for falling particles needs contactless alternatives to vibration-sensitive, batch-limited scales and costly specialized hardware. The paper introduces Doppio, a coffee-grinding video dataset with per-frame weight annotations, and evaluates spatial and spatiotemporal models. Deep-learning vision models estimate cumulative falling-particle weight, with a ResNet-34 TCN achieving a MAE of 0.18 on full test sequences.
Problem
Continuous or contactless powder-mass measurement remains underserved because scales face vibration and batch-processing limitations, while specialized alternatives are costly and complex.
Method
The paper constructs Doppio from coffee-grinding videos with per-frame weight annotations and evaluates time-based, feed-forward, GRU, and TCN models across backbone architectures.
Results
A ResNet-34 TCN achieves a MAE of 0.18 on the full test sequence, while analysis reports architecture-dependent accuracy and computational trade-offs.
Takeaways & Limitations
The work provides an accessible dataset and testbed for real-time, vision-based mass estimation of falling particles and reports generalization to unseen bean types.
Takeaways & Limitations
The models constrain outputs to positive weight estimates using a softplus activation because weight is assumed to increase monotonically over time.
Abstract
from arXiv · showhide
Measuring the mass of powder, including falling particles, is a common task in industrial applications. While scales are effective for static measurements, many applications require contactless sensing, where existing solutions are often costly, application-specific, and technically complex. In this work, we investigate computer vision as a practical alternative for contactless mass estimation. As an accessible real-world case study, we focus on coffee grinding and introduce \emph{Doppio}, a novel video dataset capturing videos of falling ground coffee, paired with precise, per-frame ground-truth weight measurements. To demonstrate contactless measuring, we evaluate deep learning-based approaches ranging from purely spatial feed-forward networks to recurrent spatio-temporal models. These models are analyzed with respect to their predictive accuracy and computational trade-offs. We demonstrate that deep learning-based computer vision models accurately estimate the cumulative weight of falling particles, establishing a solid foundation for future vision-based contactless measurement solutions.
1 Introduction
Precise, continuous powder-mass measurement remains difficult because scales are limited by vibration and batch operation, while specialized contactless systems are costly and complex. The paper uses coffee grinding to introduce a dataset and models for vision-based estimation of falling-particle weight.
- Mechanical vibrations and batch-only operation limit scales for continuous or contactless powder-mass measurement.
- Specialized contactless approaches such as microwave radar and X-ray increase engineering effort, technical complexity, and production cost.
- Optical sensors and computer vision offer potentially low-cost alternatives that can reduce measurement-system costs and technical complexity.
- Coffee grinding provides a challenging case study because agglomerated particles vary in speed and cause occlusions as they fall.
- Doppio introduces videos of falling ground coffee across roast types and grind settings, paired with models ranging from time-based baselines to recurrent neural networks.
2 Related Work
Prior work has used computer vision and other sensors for static analysis or mass-flow estimation, but dynamic weight quantification of falling coffee remains unaddressed. This work shifts toward spatiotemporal, real-time contactless estimation using the Doppio dataset.
- Computer-vision studies have addressed livestock, food-processing geometry, and coffee granularity, but not dynamic coffee mass quantification.
- Existing particle mass-flow systems commonly use photoreceptors, lasers, capacitance, X-ray, or microwave Doppler sensing rather than RGB imagery.
- The paper extends static, categorized, and volumetric coffee analysis toward spatiotemporal understanding for efficient, real-time contactless weight estimation.
3 The Doppio Dataset
Doppio pairs videos of falling coffee with scale-derived, per-frame cumulative-weight annotations collected across grind sizes and bean types. The dataset pipeline corrects display-reading and timing issues, then supports sequence-level and frame-level evaluation.
- 3.1 Data Acquisition: Doppio records falling coffee videos together with cumulative weight read from the grinder’s integrated scale.
- 3.1 Data Acquisition: The dataset spans 25 grind sizes and 3 coffee-bean types, with continuous sequences yielding about 32 g of ground coffee.
- 3.1 Data Acquisition: ArUco markers maintain spatial calibration across the floor, camera, grinder, lighting, and tripods during recording.
- 3.2 Postprocessing and Annotation: The annotation pipeline crops coffee and the scale display, applies OCR, filters and fills display-reading errors, and compensates for per-sequence measurement delay.
- 3.2 Postprocessing and Annotation: A seven-frame moving average smooths scale measurements, while the dataset is divided into 131 training, 13 validation, and 75 test sequences.
- 3.3 Dataset Splits: Evaluation includes start, mid, end, and full subsequences to compare model behavior across changing grinding dynamics.
- 3.4 Evaluation Metrics: MAE measures final sequence discrepancy normalized per doppio, while per-frame error measures the discrepancy throughout variable-length sequences.
4 Methods for Contactless Weight Estimation
The paper evaluates contactless weight estimation using models that range from a calibrated time-based baseline to feed-forward and recurrent vision architectures. Video-frame predictions estimate per-step weight differences, which are accumulated to recover cumulative particle weight over time.
- The evaluation spans models from a time-based baseline and linear regression to deep spatio-temporal architectures.
- Time-Based Baseline: The time-based baseline calibrates a linear regression model for each bean-type and grind-size combination.It tests whether a constant fall rate can estimate weight without visual inputs.
- Each video frame is processed to estimate the weight difference caused by visible coffee, and these differences are cumulatively summed over the sequence.
- A softplus activation restricts model outputs to positive weight estimates because cumulative weight increases monotonically.
- Vision-Based Models: Vision models use either a single RGB frame or the current frame concatenated with its pixel-wise difference from the previous frame.These inputs have channel dimensions C = 3 and C = 6, respectively.
- Vision-Based Models: The feed-forward model predicts current-frame weight differences without temporal information, while a recurrent architecture tracks particles across time.The recurrent design addresses errors from particles appearing in multiple frames and captures long-term temporal dependencies.
5 Experiments
The experiments compare training choices, architectures, inputs, and computational costs for contactless weight estimation on Doppio. Temporal models with larger ResNet backbones achieve the strongest accuracy, while lightweight MobileNet-based models offer lower computational cost.
- Training Setup: The combined loss supervises both final accumulated weight and intermediate per-frame differences to enforce correct temporal dynamics.The final loss combines sequence-level and difference losses using λ, with smooth L1 distance used throughout.
- Design Choices: Frame-difference inputs improve feed-forward accuracy for nearly all tested backbones, whereas excluding them significantly improves MobileNet-V3-L TCN accuracy.The subsequent GRU and TCN architectures therefore use the simpler current-frame inputs.
- Design Choices: λ = 10 performs best for feed-forward networks, while λ = 100 performs best for GRU- and TCN-based architectures.These architecture-specific loss-weighting factors are used in subsequent experiments.
- Main Results: All proposed vision-based architectures outperform the Time baseline across sub-sequences and metrics, with MobileNet-V3-S achieving the highest overall FFN accuracy.MobileNet-V3-L and ResNet-34 FFNs have nearly identical accuracies, so the FFN margins are small.
- Main Results: Temporal models favor larger backbones: ResNet models outperform MobileNet counterparts, and ResNet backbones perform better in GRUs and TCNs than in FFNs.The authors attribute weaker MobileNet GRU performance to aggressive feature reduction and coarse outputs lacking fine-grained details.
- Main Results: 0.18 MAE is achieved by the TCN with a ResNet-34 backbone on the full test sequence, but this model requires substantially more computation than lightweight backbones.Grad-CAM generally focuses on falling coffee, although one sample shows attention concentrated on larger lumps while ignoring smaller ones.
6 Conclusion
The paper introduces Doppio for contactless weight estimation of falling coffee particles and evaluates feed-forward, recurrent, and temporal architectures across accuracy and computational trade-offs.
- Doppio contains precise, per-frame ground-truth weight measurements of falling coffee grounds across grind sizes and bean types.
- The evaluation covers feed-forward networks, GRUs, TCNs, and backbones ranging from lightweight MobileNet-V3 variants to ResNets.
- Highly compressed models are well-suited for efficient feed-forward processing, whereas heavier convolutional networks are needed to build meaningful temporal contexts in GRUs.
- The models demonstrate generalization to unseen bean types.
- Doppio provides an accessible dataset and testbed for robust, real-time, vision-based mass estimation of falling particles.
Supplementary Material
The supplementary material reports full results for models trained with alternative loss weightings on the Doppio test set.
- Table A.1 compares λ = 10 and λ = 100 using MAE and MAE□ for full sequences on Doppio test.
A Loss Weighting
The loss-weighting analysis compares λ = 10 and λ = 100 across feed-forward, GRU, and TCN architectures using Doppio test metrics.
- λ = 10 performs best for the feed-forward network in terms of MAE.
- λ = 100 performs best for GRU- and TCN-based architectures in terms of MAE.
- For GRU- and TCN-based architectures, MAE□ is only slightly smaller for λ = 10.
B Distance Measures
The distance-measure analysis finds smooth L1 to be the best-performing choice for the loss terms in this setting.
- Smooth L1 performs best among the evaluated distance measures.MAE-based distance functions outperform MSE-based ones because predicted weights per step are below 1 g, making the quadratic penalty reduce their magnitude.