Roadmap

This page tracks the next scientific, technical, and editorial milestones for AstroVision. The purpose is not to list every small task, but to make the direction of the project clear to readers, supervisors, collaborators, and future reviewers.

Current phase

AstroVision is in a consolidation phase: the main architecture exists, the strongest result is the late-fusion multimodal model, and the next step is to strengthen reproducibility, uncertainty analysis, and manuscript readiness.

---

Roadmap overview

Phase	Focus	Status
v0.1	Baseline CNN, initial dataset preparation, first morphology classification experiments	Complete
v0.2	EfficientNet and DINOv2 comparison, stronger benchmark structure	Complete / polishing
v0.3	Late-fusion multimodal model and physical metadata integration	Active
v0.4	Uncertainty, calibration, Grad-CAM, and morphometry consolidation	Active
v0.5	Manuscript-ready figures, claims register, and reproducibility audit	Planned
v1.0	Public research release with stable documentation and citation metadata	Planned

---

Scientific priorities

1. Consolidate the late-fusion result

The late-fusion multimodal model reports 86.3% balanced accuracy on its own test split. It stays the flagship model but is not a clean modality gain: the split differs from the image-only models (best: DINOv2 fine-tuned, 83.9%), the ablation differences are below 0.4 points, and a leak through the fine-tuned backbone cannot be ruled out (see Evaluation Protocol). It needs to be documented with full context.

Required next steps:

• Verify final train/validation/test split.
• Export final confusion matrix.
• Export per-class precision, recall, and F1.
• Document whether metadata improves specific classes more than others.
• Ablation table (DINOv2 only, + morphometry, + SDSS photometry, + both): published in Key Results.
• Add a metadata-only configuration (morphometry + photometry, no image features) to the ablation: the current ablation always keeps DINOv2.
• Add the final claim to the Scientific Claims Register.

2. Make calibration findings explicit

The uncertainty page should not only report successful techniques. It should also preserve the finding that Temperature Scaling did not improve the late-fusion model in the current run.

Required next steps:

• Add reliability diagram.
• Add ECE table before/after calibration.
• Explain why Temperature Scaling may fail for this setting.
• Compare MC Dropout uncertainty against correctness.
• Connect conformal prediction to scientifically cautious classification.

3. Strengthen the physical interpretation

AstroVision should avoid becoming a pure computer-vision benchmark. The strongest version of the project explains what the model appears to learn about galaxy structure.

Required next steps:

• Link Grad-CAM examples to physical regions of galaxies.
• Compare morphology classes using morphometry descriptors.
• Add representative success and failure cases.
• Identify which classes are physically ambiguous rather than merely computationally hard.
• Clarify why disturbed galaxies remain difficult.

---

Technical priorities

Priority	Task	Target page
P1	Add full evaluation protocol	Evaluation Protocol
P1	Complete benchmarking table	Benchmarking
P1	Finalize late-fusion architecture description	Late-Fusion
P1	Document disturbed-class limitation	Limitations
P2	Add segmentation quality examples	Segmentation
P2	Add morphometry feature definitions	Morphometry
P2	Add API usage examples	API Reference
P3	Add automated report generation	Guides / API

---

Manuscript preparation

AstroVision is a natural candidate for a compact methods-and-results manuscript if the following pieces are completed.

Minimum manuscript-ready package

• Abstract-level project framing.
• Dataset and preprocessing section.
• Model architecture figure.
• Benchmark table.
• Confusion matrix.
• Calibration / uncertainty figure.
• Grad-CAM composite figure.
• Morphometry / physical interpretation figure.
• Limitations section.
• Reproducibility statement.
• Citation / release metadata.

Suggested manuscript narrative

1. Galaxy morphology classification is visually rich but physically ambiguous.
2. Image-only deep learning provides strong baselines.
3. Foundation vision features help, but do not automatically solve all class ambiguities.
4. Late-fusion improves the scientific framing by combining visual representation with physically interpretable descriptors.
5. Uncertainty and conformal prediction are necessary because morphology labels can be ambiguous.
6. Failure cases, especially disturbed galaxies, are scientifically informative rather than merely technical errors.

---

Documentation milestones

Page	Status goal
Overview	Replace scaffold bullets with a narrative project introduction
Key Results	Add final metric cards and caveat box
Evaluation Protocol	Complete and keep stable
Benchmarking	Add final comparison table
Uncertainty	Highlight calibration outcome
Grad-CAM	Add figure gallery and interpretation rules
Physical Interpretation	Connect predictions to galaxy structure
Limitations	Expand disturbed-galaxy analysis
Citing	Update once DOI / release exists

---

Release checklist

Before tagging a public release:

• Every sidebar page builds without broken links.
• Every major claim has a source figure or table.
• Every result page states the dataset/split/model version.
• The late-fusion result is framed with caveats.
• The disturbed-class limitation is clearly visible.
• The citation page includes release metadata.
• The documentation landing page links to AstroVision key results.
• The repository has a reproducible environment file.
• The final figures have consistent style and captions.

---

Related pages

Page	Role
Key Results	Current scientific summary
Late-Fusion Multimodal	Main architecture
Evaluation Protocol	Stable evaluation rules
Uncertainty	Calibration and confidence
Limitations	Failure modes and caveats
Citing	Citation metadata