Glass-Filling with Prioritization

• Glass-filling with prioritization is a concept that applies selective ranking to ambiguous states across computer vision, robotic perception, glass science, and generative inverse design.
• Techniques include class-weighted loss, adversarial transfer learning, robust depth-recovery, and topological constraint prioritization to optimize performance and resource allocation.
• Prioritization strategies lead to practical benefits such as enhanced classification accuracy, optimized simulation resources, improved material properties, and efficient candidate selection.

Glass-filling with prioritization denotes a class of technical problems in which a filling state, a glass-composition state, or an evaluation queue is not treated uniformly, but is instead estimated, explored, or acted upon under explicit priorities. In computer vision, the term can refer to single-image filling-level classification for cups and drinking glasses, including the prioritization of specific fill classes, robustness regimes, or container types. In robotic perception, it can refer to using transparent-object depth completion to recover vessel geometry and then rank glasses by fill fraction or uncertainty. In glass science, it can refer to populating composition space with selected structural units so that the resulting network occupies a desired rigidity regime. In generative inverse design, it can refer to queue prioritization, where a design space is “filled” with evaluated candidates according to an acquisition score rather than arrival order (Modas et al., 2021, Xu et al., 2021, Bauchy, 2020, Schwarting et al., 29 Sep 2025).

1. Scope and technical meanings

The phrase spans several research traditions rather than a single canonical formalism. This suggests an umbrella usage centered on selective filling: selective labeling of vessel contents, selective reconstruction of transparent vessels for pouring, selective occupation of composition space in glass science, and selective advancement of generated candidates in active-learning workflows.

Technical setting	“Glass” denotes	Prioritized quantity
Single-image classification	cups and drinking glasses	filling level, robustness, container types
Transparent-object perception	transparent vessels in cluttered RGB-D scenes	fill fraction, target sequence, confidence-aware ranking
Topological constraint theory	composition space and the “glass genome”	constraint density and rigidity state
Generative inverse design	design space $\Omega$	acquisition score and queue order

Across these settings, prioritization is not merely an ordering heuristic. It changes the effective objective. In the fill-level classifier, it can be introduced through class weighting, inverse-frequency sampling, or emphasis on occluded and transparent cases. In transparent-vessel reconstruction, it can determine which vessel is filled first or which estimate is trusted. In topological constraint theory, it determines which structural units and compositions are explored first. In queue-prioritized inverse design, it governs which candidates consume simulation or experimental resources first.

2. Single-image filling-level classification

A direct meaning of glass-filling with prioritization is the estimation of content level in a cup or a drinking glass from a single RGB image. The task is formulated as single-image multi-class classification with input $I \in [0,1]^{H \times W \times C}$ and output label $y \in \{0\%, 50\%, 90\%, \text{unknown}\}$, where $0\%$ denotes empty, $50\%$ half-full, $90\%$ almost full, and unknown denotes opaque or translucent containers or content where liquid level is not visually measurable (Modas et al., 2021).

The target domain is the CORSMAL Containers Manipulation dataset restricted to 4 cups and 4 drinking glasses. From CCM videos, frames are automatically sampled where pouring is finished, the container region is cropped using Mask R-CNN, and each crop is manually verified and annotated. The resulting C-CCM dataset contains 10,269 images. Each image is annotated with filling type, filling level, hand occlusion, and container transparency. Three train/test splits, S1, S2, and S3, are used to study shape generalization. The setup is explicitly designed to stress transparency, shape variability, partial occlusion by hands, and small-sample overfitting.

The classifier is ResNet-18 with a final fully connected layer producing a 4-way softmax. Supervised training uses cross-entropy loss,

$$\mathcal{L}_{\text{cls}}(I, y \mid \theta) = -\sum_{k=1}^{4} y_k \log p_k .$$

Class imbalance is handled by probability-proportional-to-inverse-frequency sampling, so each batch is formed by randomly sampling images with probabilities inversely proportional to the number of images per class.

The central methodological result is a transfer-learning strategy in which adversarial training is performed on the source domain, ImageNet, and only then refined on C-CCM. Six strategies are evaluated: ST, AT, ST$\rightarrow$FT, ST$\rightarrow$AFT, AT$\rightarrow$FT, and AT$I \in [0,1]^{H \times W \times C}$0AFT. The best-performing and most important strategy is AT$I \in [0,1]^{H \times W \times C}$1FT: pre-train on ImageNet with adversarial training, freeze the first residual block group, and fine-tune the remaining layers on clean C-CCM data. Best performance occurs at $I \in [0,1]^{H \times W \times C}$2, with split accuracies of 78.34\% on S1, 65.63\% on S2, and 82.32\% on S3 for the frozen-layer sensitivity study. For AT$I \in [0,1]^{H \times W \times C}$3FT, the best source-domain perturbation strengths are $I \in [0,1]^{H \times W \times C}$4 for S1 with 80.97\%, $I \in [0,1]^{H \times W \times C}$5 for S2 with 73.27\%, and $I \in [0,1]^{H \times W \times C}$6 for S3 with 88.23\%.

Per-container behavior is highly nonuniform. On S1, the champagne flute is difficult for ST and AT, both at about 46\% accuracy, whereas AT$I \in [0,1]^{H \times W \times C}$7FT improves performance by about 30 percentage points over ST. On S3, the green glass improves from about 66\% under ST to about 10 percentage points higher under AT$I \in [0,1]^{H \times W \times C}$8FT. The red cup is the most extreme case: ST accuracy is 0.005\%, all transfer-learning techniques surpass 90\%, and AT$I \in [0,1]^{H \times W \times C}$9FT achieves ~99.5\% accuracy. The reported interpretation is that adversarial training on ImageNet followed by fine-tuning on C-CCM consistently yields the highest per-container accuracy and best generalization to unseen containers, particularly those differing in shape or transparency.

The paper itself does not explicitly discuss prioritization. It does, however, provide an explicit extension path. Class-weighted cross-entropy can be introduced as

$y \in \{0\%, 50\%, 90\%, \text{unknown}\}$0

so that $y \in \{0\%, 50\%, 90\%, \text{unknown}\}$1 versus $y \in \{0\%, 50\%, 90\%, \text{unknown}\}$2, $y \in \{0\%, 50\%, 90\%, \text{unknown}\}$3, or unknown can be emphasized according to application-specific cost. The same logic extends to importance weighting by container type, transparency type, or hand occlusion: $y \in \{0\%, 50\%, 90\%, \text{unknown}\}$4 Because C-CCM includes transparency and hand-occlusion annotations, robustness can be prioritized for transparent containers, transparent liquids, or occluded images.

3. Transparent-vessel geometry and prioritized robotic filling

A second meaning of glass-filling with prioritization arises in robotic perception, where the difficulty is not only to infer a discrete fill class but to recover vessel geometry and liquid-surface cues in cluttered RGB-D scenes containing transparent objects. "Seeing Glass" introduces TranspareNet, a joint point cloud and depth completion method, and the Toronto Transparent Objects Depth Dataset (TODD), created through robot-controlled image collection and vision-based automatic annotation (Xu et al., 2021).

TranspareNet starts from the observation that commodity RGB-D sensors provide distorted rather than entirely useless depth measurements on transparent objects. The pipeline assumes an instance mask for each transparent object, de-projects masked raw depth into a sparse distorted point cloud, completes that point cloud using a GRNet-based point cloud completion module, projects the result back into the depth domain, and finally applies an RGB-D depth-completion network with EfficientNet-B4 encoder and a U-Net-style decoder modulated by SPADE. The point-cloud stage is trained with a gridding loss,

$y \in \{0\%, 50\%, 90\%, \text{unknown}\}$5

and the depth-completion stage uses a log-$y \in \{0\%, 50\%, 90\%, \text{unknown}\}$6 pair-wise loss on valid ground-truth pixels.

TODD contains 14,659 RGB-D images, with up to 3 transparent objects per scene, five background templates, six real glass vessels, and objects that are empty or filled with 5 different colored liquids. The hardware is an eye-in-hand Franka Emika Panda robot with an Intel RealSense D435i at 640×480 resolution. The annotation pipeline projects CAD meshes, using AprilTag-based pose estimation, to obtain per-object masks, ground-truth depth, and 6DoF object pose.

The reported performance is strong on both ClearGrasp and TODD. On ClearGrasp, for seen objects, TranspareNet-DC only reaches RMSE 0.011, REL 0.010, and $y \in \{0\%, 50\%, 90\%, \text{unknown}\}$7, compared with ClearGrasp at 0.041, 0.055, and 0.6943. On TODD, for the novel combined set, TranspareNet reaches RMSE = 0.0213, REL = 0.0510, and $y \in \{0\%, 50\%, 90\%, \text{unknown}\}$8. The method also improves downstream pose estimation: MaskedFusion with TranspareNet depth attains Average ADD 0.01209 and 88.14\% with ADD < 2 cm, close to the results using ground-truth depth.

The paper does not explicitly model liquid level or fluid dynamics, but the completed depth can support downstream fill-level estimation. A vessel point cloud can be obtained by de-projecting the completed depth within the object mask,

$y \in \{0\%, 50\%, 90\%, \text{unknown}\}$9

Given a vessel model and a fitted liquid plane, liquid volume and fill fraction can be estimated as

$0\%$0

For cylindrical glasses, this reduces to $0\%$1.

Prioritization then enters at the task level. The synthesized workflow defines scores such as fill-the-emptiest-first, $0\%$2; target-fill ranking, $0\%$3; accessibility-weighted prioritization,

$0\%$4

and confidence-aware prioritization,

$0\%$5

These are downstream constructs rather than native TranspareNet outputs. The paper’s direct contribution is the perception substrate: accurate depth for transparent vessels in clutter, including partially filled fluid contents within the vessels.

4. Topological constraint theory and compositional filling of glass space

In glass science, glass-filling with prioritization has a distinct meaning. It refers to populating composition space with selected structural units so that the resulting atomic network falls into a targeted rigidity regime. "Topological Constraint Theory and Rigidity of Glasses" presents the formal basis: a disordered atomic glass network is mapped to a mechanical truss, and the controlling descriptor is the number of topological constraints per atom (Bauchy, 2020).

For a 3D truss of $0\%$6 nodes and $0\%$7 mechanical constraints, Maxwell counting gives

$0\%$8

and, for macroscopic atomic networks, the floppy modes per atom satisfy

$0\%$9

where $50\%$0. The rigidity classes are then: flexible if $50\%$1, isostatic if $50\%$2, and stressed-rigid if $50\%$3. In simple covalent networks with all constraints intact and no 1-fold coordinated atoms, an atom of coordination $50\%$4 contributes

$50\%$5

and the chapter gives

$50\%$6

The isostatic condition yields the canonical coordination

$50\%$7

This counting scheme is extended to one-fold species, temperature-dependent constraints, and pressure-dependent constraints. With a fraction $50\%$8 of 1-fold coordinated atoms,

$50\%$9

Mauro and co-workers introduce onset temperatures $90\%$0 for each constraint so that angular constraints typically break earlier than radial constraints. Bauchy and Micoulaut show that pressure can change coordination and thus $90\%$1, inducing rigidity transitions. Boolchand et al. argue for an intermediate phase, an extended composition range in which the network self-organizes into an isostatic, stress-free state bounded by a flexible-to-rigid transition and an unstressed-to-stressed transition.

This framework turns prioritization into a topological design rule. The central prioritization axis is constraint density / rigidity state. Structural units such as SiO$90\%$2, BO$90\%$3, PO$90\%$4, GeSe$90\%$5, AsSe$90\%$6, bridging oxygens, and non-bridging oxygens are not equivalent: each contributes a characteristic pattern of bond-stretching and bond-bending constraints, and therefore shifts the network toward flexibility, isostaticity, or stressed rigidity.

5. Property-targeted prioritization in glass composition design

Once the constraint-counting framework is established, prioritization becomes property-directed. The chapter explicitly connects rigidity to glass-forming ability, hardness, fracture toughness, viscosity, fragility, glass transition temperature, and dissolution kinetics (Bauchy, 2020).

Canonical examples show how isostatic compositions become priority targets. In Ge$90\%$7Se$90\%$8,

$90\%$9

so $$\mathcal{L}_{\text{cls}}(I, y \mid \theta) = -\sum_{k=1}^{4} y_k \log p_k .$$0 gives $$\mathcal{L}_{\text{cls}}(I, y \mid \theta) = -\sum_{k=1}^{4} y_k \log p_k .$$1, corresponding to GeSe$$\mathcal{L}_{\text{cls}}(I, y \mid \theta) = -\sum_{k=1}^{4} y_k \log p_k .$$2. Experimentally, these glasses show highest glass-forming ability around $$\mathcal{L}_{\text{cls}}(I, y \mid \theta) = -\sum_{k=1}^{4} y_k \log p_k .$$3 Ge, minimum non-reversible enthalpy at $$\mathcal{L}_{\text{cls}}(I, y \mid \theta) = -\sum_{k=1}^{4} y_k \log p_k .$$4, and maximum space-filling around $$\mathcal{L}_{\text{cls}}(I, y \mid \theta) = -\sum_{k=1}^{4} y_k \log p_k .$$5. In As$$\mathcal{L}_{\text{cls}}(I, y \mid \theta) = -\sum_{k=1}^{4} y_k \log p_k .$$6Se$$\mathcal{L}_{\text{cls}}(I, y \mid \theta) = -\sum_{k=1}^{4} y_k \log p_k .$$7,

$$\mathcal{L}_{\text{cls}}(I, y \mid \theta) = -\sum_{k=1}^{4} y_k \log p_k .$$8

so $$\mathcal{L}_{\text{cls}}(I, y \mid \theta) = -\sum_{k=1}^{4} y_k \log p_k .$$9, corresponding to As$\rightarrow$0Se$\rightarrow$1, again coinciding with optimal glass-forming ability, minimal enthalpy relaxation, and maximal space-filling. Silica, boron oxide, and phosphorus pentoxide are each shown to be isostatic once the appropriate constraint assumptions are made, including the thermally flexible Si–O–Si angle in silica. In sodium silicate,

$\rightarrow$2

and isostaticity occurs at $\rightarrow$3, where experiments show minimum non-reversible enthalpy around 20 mol\% Na$\rightarrow$4O.

Phillips’ original hypothesis follows directly: the best glass formers are isostatic. Flexible networks contain floppy modes that facilitate atomic rearrangements, relaxation, and crystallization; stressed-rigid networks contain redundant constraints and internal eigenstress; isostatic networks have neither floppy modes nor eigenstress. From a prioritization perspective, the intermediate phase and the isostatic window become hot spots for formulations seeking robust glass-forming ability, minimized relaxation, and balanced mechanical behavior.

The chapter also reviews analytic or semi-analytic property models. For hardness,

$\rightarrow$5

with $\rightarrow$6. It is further reported that hardness often correlates more strongly with the number of angular constraints per atom than with total $\rightarrow$7. Fracture toughness often peaks near $\rightarrow$8, as seen in densified sodium silicate glasses, calcium-silicate-hydrate phases, and Ge–Se glasses. For dissolution kinetics,

$\rightarrow$9

with $\rightarrow$0 kJ/mol. Higher $\rightarrow$1 therefore implies slower dissolution, whereas lower $\rightarrow$2 implies faster dissolution.

These relations generate a topological map for prioritized compositional exploration. To design hard or chemically durable glasses, one prioritizes structural motifs that increase bond-bending constraints and total $\rightarrow$3. To design tough glasses, one prioritizes compositions in or near the isostatic window. To design processable or reactive glasses, one prioritizes more flexible or mildly rigid networks. The chapter nonetheless emphasizes that topological constraint theory is a first-order filter: it intentionally filters out exact bond lengths, local chemical heterogeneity, medium-range order, and dynamic heterogeneity, and therefore should be refined by molecular dynamics, NMR, diffraction, Raman, and related structural analyses.

6. Queue prioritization in generative inverse design

A different but conceptually related use of glass-filling with prioritization appears in generative materials discovery. "Steering an Active Learning Workflow Towards Novel Materials Discovery via Queue Prioritization" treats the design space as a “glass” to be filled with evaluated candidates, while an active-learning model determines which candidates enter expensive evaluation first (Schwarting et al., 29 Sep 2025).

The workflow has four main components. A pre-trained generator $\rightarrow$4 produces candidates that are appended to a generation queue $\rightarrow$5. A surrogate model $\rightarrow$6, implemented as an XGBoost regressor on 38 RDKit features, predicts an expensive property, the MD-based internal strain $\rightarrow$7, and an uncertainty estimate $\rightarrow$8. An acquisition function $\rightarrow$9 reorders candidates into an updated or prioritized queue $\rightarrow$0. High-priority candidates are then assembled and simulated with LAMMPS, and the resulting labels are used to retrain the surrogate and fine-tune the generator on high-performing candidates.

The surrogate minimizes

$\rightarrow$1

Acquisition can be exploit-only, $\rightarrow$2; uncertainty-only, $\rightarrow$3; or Upper Confidence Bound-style, for example

$\rightarrow$4

The workflow also supports linear combinations of stability, synthesizability, and HMOF similarity: $\rightarrow$5

$\rightarrow$6

$\rightarrow$7

The reported application is carbon-capture MOF molecular design. The expensive target metric is internal strain, with MOF stability corresponding to $\rightarrow$8 below about 0.25. The generator is DiffLinker, fine-tuned on HMOF linkers. Over runs with exactly 1000 novel candidates, the baseline workflow without active learning generates an average of 281 high-performing candidates, whereas prioritization with active learning generates an average of 604 high-performing candidates. The paper further reports that weak UCB yields the lowest RMSE on a hold-out set and the greatest number of stable MOFs, explore-only acquisition yields poorer training and fewer stable MOFs, and active-learning queue prioritization mitigates generative model decay when the generator is fine-tuned on large fractions of its own outputs.

The compute profile is also explicit. MD simulation dominates resource use, consuming 48 CPUs, 98.5\% time utilization, and 94.6\% of total CPU utilization. The AL prioritization worker runs on 1 CPU, at about 28.9\% of that CPU’s time and about 0.6\% of total CPU resources. Each prioritization pass takes about 16 seconds on one CPU, and the AL worker is triggered about 120–130 times during 1000-candidate runs. In this setting, prioritization is a queueing mechanism with direct scientific consequences: it prevents resources from being spent on nonsensical candidates and changes which candidates become training signal for subsequent generative updates.

7. Limits, misconceptions, and unifying structure

Several misconceptions recur across these usages. In the fill-level classification setting, adversarial training is adversarial example training in the sense of Madry et al. using PGD, with $\rightarrow$9-bounded perturbations and 10-step PGD; it is not combined with any domain-adversarial objective or GAN (Modas et al., 2021). In the transparent-object setting, TranspareNet does not explicitly model liquid level or fluid dynamics; it reconstructs geometry, while liquid-surface fitting, fill-fraction estimation, and task-level prioritization are downstream additions (Xu et al., 2021). In topological constraint theory, $I \in [0,1]^{H \times W \times C}$00 is a powerful descriptor but not a complete structural theory; the chapter explicitly notes that TCT filters out bond lengths, local chemical heterogeneity, medium-range order, and dynamic heterogeneity, and therefore should be complemented by structural analysis and simulation (Bauchy, 2020). In queue-prioritized inverse design, the benefit depends on surrogate quality and hyperparameter choices such as the exploration parameter in UCB and the fine-tuning fraction used for the generator (Schwarting et al., 29 Sep 2025).

A broader unifying structure nevertheless emerges. In all four settings, a large ambiguous space is reduced by a priority mechanism. In image-based glass filling, robust source-domain features are used so that the classifier does not overfit to specific CCM training shapes, colors, or backgrounds. In transparent-vessel perception, raw but distorted sensor depth is not discarded; it is used as a prior for recovering geometry in cluttered scenes. In glass composition design, structural complexity is reduced to topological descriptors such as $I \in [0,1]^{H \times W \times C}$01, $I \in [0,1]^{H \times W \times C}$02, and $I \in [0,1]^{H \times W \times C}$03, and compositions are prioritized according to rigidity regime. In inverse design, a candidate stream is reordered by acquisition scores so that simulation and fine-tuning are concentrated on promising regions of design space.

This suggests that glass-filling with prioritization is best understood not as a single named algorithm, but as a recurring methodological pattern: define a space to be filled, define a scarcity or ambiguity that makes uniform treatment suboptimal, and then introduce a principled ranking variable—class weight, fill fraction, rigidity state, uncertainty, predicted performance, or queue score—to determine what should be estimated, explored, or acted on first.