MSD Dataset: Polysemous Research Benchmarks

• MSD Dataset is a polysemous collection of benchmark datasets defined by distinct domains such as architecture, medicine, facial recognition, natural language, and agriculture.
• Each dataset employs rigorous annotation pipelines and specialized evaluation metrics like Dice similarity and IoU to ensure domain-specific performance assessment.
• These resources drive advancements in machine learning by challenging models to adapt to complex, multi-domain scenarios with real-world applicability.

The acronym "MSD Dataset" appears across multiple research domains, denoting at least five distinct datasets: (1) Modified Swiss Dwellings (floor plans), (2) Medical Segmentation Decathlon (medical imaging), (3) Masked Student Dataset of Expressions (facial expression recognition under mask occlusion), (4) Multilingual Simile Dialogue (figurative language in dialogue), and (5) Maize Seedling Detection Dataset (agriculture/plant phenotyping). Each instantiation is prominent within its respective community and serves as a specialized benchmark to advance domain-specific machine perception, reasoning, or decision tasks.

1. Modified Swiss Dwellings (MSD) for Floor Plan Generation

Scope and Representation:

The Modified Swiss Dwellings (MSD) dataset is a large-scale collection of floor plans from Switzerland, focusing on the generation and semantic understanding of medium- and large-scale (multi-apartment) residential building complexes. It comprises 5,372 raster-vector plans representing over 18,900 apartments, with each plan annotated by: (1) a room-type mask, (2) real-world X/Y coordinate maps, (3) a binary structural-wall mask, and (4) fine-grained vector and graph data detailing room and door polygons as well as connectivity and zoning attributes.

Annotation Pipeline and Composition:

Vector-graphics CAD plans serve as input; non-residential or duplicate plans are filtered, and rooms and doors are algorithmically extracted. Zoning graphs (private, public, service, outdoor) and multiple connectivity graphs (room, zoning) are derived via polygon proximity and adjacency heuristics. Structural masks are generated via morphological thinning and quantile-based wall selection (Engelenburg et al., 2024).

Statistical Observations:

MSD plans are highly complex: average corners per room (8.68 vs. ~5.0 in RPLAN), rooms per unit (mean 8.75), and high non-Manhattan (irregular) room prevalence (~51%). Entropy over room graphs (Hg = 8.02 bits) surpasses previous large-scale plan datasets (e.g., RPLAN: 4.56). These properties make it a uniquely challenging benchmark for machine-learning models aiming at topology-aware, geometry-constrained, and semantically rich floor plan generation.

2. MSD in Medical Imaging: Medical Segmentation Decathlon

Nature and Tasks:

The Medical Segmentation Decathlon (MSD) is an aggregation of 10 anatomically and modality-diverse 3D medical image datasets supporting robust, domain-generalized segmentation benchmarks. Key tasks include brain tumor, heart, liver, hippocampus, prostate, lung, pancreas, hepatic vessel, spleen, and colon segmentation (Liu et al., 2020).

Benchmarking and Architectures:

MSD catalyzed the development of generalizable, universal segmentation networks, exemplified by FIRENet—a 3D encoder–decoder with a fabric-style bottleneck for multi-scale feature aggregation and transfer learning across medical domains. Preprocessing protocols include isotropic 1mm³ resampling, cropping to fit GPU memory constraints (~180³ voxels), and per-dataset normalization, supporting unified training and evaluation.

Performance Metrics:

Dice similarity coefficient (DSC) is the primary evaluation metric, as volume overlap is central in medical segmentation. Multi-decoder architectures allow distinct label sets per sub-task, enabling efficient, task-specific output while sharing backbone representations.

3. Masked Student Dataset of Expressions (MSD-E)

Dataset Overview:

MSD-E is a facial expression recognition (FER) resource tailored for analyzing the impact of mask occlusion. It consists of 1,960 real-world images (964 non-masked, 996 masked) from 142 Indian participants, each labeled with one of seven basic expressions. The dataset reflects diverse mask types, occlusion patterns, and demographic origin (Sola et al., 2023).

Experimental Paradigms and Results:

ResNet-18 baselines demonstrate a substantial accuracy drop in masked FER (non-masked: 65.35%, masked: 45.85%). Jointly training on both types degrades the non-masked recognition performance. Advanced strategies, such as contrastive learning and knowledge distillation, partially recover accuracy by encouraging representation sharing across occlusion states without increasing inference cost. State-of-the-art occlusion-robust FER models (e.g., SCAN) outperform ResNet-18, albeit with much larger capacity.

Application and Generalization:

MSD-E reveals that synthetic mask occlusion datasets do not generalize seamlessly to real mask conditions, emphasizing its utility as a benchmark for developing mask-robust FER systems.

4. Multilingual Simile Dialogue Dataset (MSD)

Construction and Annotation:

The MSD Simile Dialogue corpus is the largest manually annotated resource for figurative (simile) language in dialogue, containing ∼20,000 annotated dialogues (Chinese/English). Data is drawn from Reddit (English) and Weibo (Chinese), filtered using comparator keywords ("like", "as...as", and Chinese equivalents), with detailed labeling of tenor, vehicle, comparator, and (for English) shared property (Ma et al., 2023).

Task Structure:

MSD enables the study of (1) simile recognition (distinguishing figurative from literal usage), (2) property and vehicle interpretation (multi-choice fill-ins), and (3) retrieval/generation tasks in dialogue context, challenging models to exploit discursiveness, cross-utterance linkage, and non-nominal simile structure.

Empirical Findings:

Baseline results reveal that simile-aware tasks are substantially harder than literal-only dialogue tasks: BERT achieves F1 ≈ 0.70 in recognition, hit@1 ≈ 0.56 in property interpretation, and response retrieval on MSD similes is 40–50% lower than on general dialogue retrieval, indicating significant headroom for model development.

5. Maize Seedling Detection Dataset (MSDD)

Acquisition and Structure:

MSDD is a high-resolution RGB aerial imagery dataset from the University of Missouri, designed for maize stand counting in precision agriculture. The dataset spans four growing seasons and contains 3,146 image fragments with 163,921 annotated maize seedlings, demarcated as single, double, or triple stands. Annotation protocols combine manual bounding-box delineation, semi-automatic YOLO-assisted proposal correction, and geometric homography-based propagation across video frames (Kharismawati et al., 18 Sep 2025).

Distribution and Class Imbalance:

Class representations are heavily imbalanced: single (92.47%), double (6.07%), triple (1.45%) per global distribution. This skews detection error towards rare classes; YOLO variants and Faster-RCNN perform optimally on singles, while doubles and triples remain challenging ([email protected] ≈ 0.233 for doubles, ≈0.313 for triples at best).

Benchmarking Protocols:

Evaluation uses IoU-based matching, precision/recall/F1 and mean average precision ([email protected], 0.50:0.95). YOLO-11x reaches inference speeds of 27 fps with acceptable accuracy for most application contexts.

Challenges and Future Directions:

Persistent limitations include annotation ambiguity (overlaps, labelling errors), perspective and illumination variability, and a need for rare-class augmentation. Prospective directions address instance segmentation, oblique-view diversity, and other crop/weed detection extensions.

6. Methodological and Naming Ambiguity in "MSD Dataset"

The acronym "MSD Dataset" has no universal meaning; each usage is officially named and domain-local (e.g., Modified Swiss Dwellings for architecture, Medical Segmentation Decathlon for imaging, Masked Student Dataset for FER, Maize Seedling Detection Dataset for plant phenotyping, Multilingual Simile Dialogue for NLP). Researchers must disambiguate by context—task domain, citation, or metadata.

MSD Usage	Domain	Reference
Modified Swiss Dwellings	Floor plan generation	(Engelenburg et al., 2024, Kuhn, 2023)
Medical Segmentation Decathlon	3D medical imaging	(Liu et al., 2020)
Masked Student Dataset (MSD-E)	FER / occlusion analysis	(Sola et al., 2023)
Multilingual Simile Dialogue	NLP / figurative lang.	(Ma et al., 2023)
Maize Seedling Detection	Plant phenotyping	(Kharismawati et al., 18 Sep 2025)

This tabulation clarifies that "MSD dataset" is polysemous; precise identification is necessary for rigorous comparison.

7. Conclusion

"MSD Dataset" denotes distinct, high-impact resources in several research areas. Each incarnation provides rigorously annotated, task-specific data, and has established itself as a benchmark within its respective subfield. Researchers are advised to specify the dataset context (domain, full name, citation) to avoid ambiguity. Across domains, the commonality is that each "MSD" delivers a challenging, large-scale, and richly annotated testbed—whether for spatial reasoning, vision, natural language understanding, facial recognition under occlusion, or agricultural scene analysis. The clarity of labelling, comprehensiveness of modalities, and open access have allowed these datasets to drive significant architectural and methodological advances. Future dataset releases are likely to continue the polysemy of the "MSD" acronym, reinforcing the need for precise referencing.