The five dimensions a vision-course robustness project is judged on. Each pairs the signals that mark strong work with the ones that weaken it. What the project must contain is a separate question, covered in the Project Requirements for Vision Course guide.
The Five Dimensions
Grading weighs all five together; a project strong on rigor but silent on visualization is incomplete, and vice versa.
Criterion 1
Study Design & Selection
Whether the dataset, tasks, and distortions were chosen deliberately and justified: ground truth is available, the tasks span low- and high-level vision, and the distortions are realistic for the domain.
Strong
Tasks span classical and deep methods.
Domain-appropriate distortions, justified with links.
Fixed, seeded dataset subset.
Weak
Three variants of one task.
Distortions unrelated to the domain.
Choices asserted, never justified.
Criterion 2
Measurement Rigor
Whether performance is measured correctly: a clean baseline against ground truth, task-appropriate metrics, severity calibrated to SNR/PSNR, and results broken down per class and per severity level.
Strong
Correct metric per task (mAP, mIoU, match ratio).
Degradation plotted against SNR in dB.
Per-class, per-SNR tables.
Weak
Demo outputs instead of metrics.
One aggregate number hiding class collapse.
Severity as an arbitrary knob value.
Criterion 3
Depth of Exploration
How thoroughly the space is covered: multiple tasks, distortions, and severity levels, and both recovery strategies explored rather than a single point measured. Negative and surprising results are kept, not hidden.
Strong
Dense severity sweeps, not two points.
Both enhancement and fine-tuning tried.
Failures reported honestly.
Weak
A single distortion at one level.
Only one recovery strategy attempted.
Cherry-picked best-case images.
Criterion 4
Recovery Analysis & Insight
Whether enhancement and fine-tuning are compared fairly on the same tasks and severities, and whether the project draws a defensible conclusion about when each strategy helps and why.
Strong
Head-to-head on identical conditions.
A clear when-does-what-win conclusion.
Links results to task abstraction level.
Weak
Strategies compared on different subsets.
Numbers with no interpretation.
Claiming recovery without a baseline delta.
Criterion 5
Visualization & Communication
How clearly the work is shown and defended. Strong projects make the finding visible: images with annotations overlaid, before/after grids across distortion and restoration, and measurement plots (per-SNR curves, per-class bars). The three panels below are the minimum visual vocabulary a report should speak.
