Release task
Inspect the task files, instructions, source material, and evaluation criteria.
Task requirements
Images load automatically. YAML loads when expanded; audio and other formats remain metadata-only.
Release task
Inspect the task files, instructions, source material, and evaluation criteria.
1 source files available in data/.
Task requirements
The only model-facing image is:
data/alloy_M17_micrograph.png
It is a deterministic synthetic etched-alloy optical micrograph at native resolution with a visible scale bar and label in the light annotation strip at the bottom. The field contains realistic acquisition and preparation defects: radial vignetting / uneven illumination with a measurable corner-minus-center dark bias, a long dark polishing scratch, isolated surface debris specks, and a zone with locally weak grain boundaries and touching grains. Per-pixel sensor/etch noise causes pore, artifact, and boundary intensity distributions to overlap; no single raw luma threshold is a complete solution.
Use the native image, not a screenshot. No other images or ground-truth masks are provided to you.
Produce a defensible quantitative metallography package: calibrate from the scale bar, define the analysis ROI, correct illumination, segment true pores while rejecting scratch/debris, separate individual grains despite weak boundaries, assign every ROI pixel to exactly one grain territory, measure per-grain area, centroid and equivalent-circle diameter (ECD), compute porosity and a planimetric-style grain density estimate, and deliver auditable masks, spatial labels, overlays, statistics, and report.
Fiji / ImageJ is available in the environment and is encouraged for interactive inspection, scale-bar measurement, background correction, and marker-controlled watershed, because this is a professional image-metrology workflow. Python image-analysis tools (scikit-image, scipy, numpy, matplotlib, Pillow, pandas) are also acceptable. Grading scores outputs, not tool choice — do not lose points for using Python if your masks and numbers are correct, and do not gain points for mentioning Fiji without delivering correct artifacts.
Calibration. Read the physical length from the scale-bar label in the image. Measure the scale-bar length in pixels on the native image. Compute um_per_px = physical_length_um / length_px. State method, measured pixel length, and resulting calibration_um_per_px. Use this calibration for all um and um2 conversions.
Analysis ROI. Restrict all quantitative analysis to the rectangular micrograph field above the light annotation strip at the bottom. The annotation strip and the scale bar inside it must be excluded from the porosity denominator and from grain/field area denominators. Determine the rectangular ROI visually and record it as analysis_roi_xywh_px = [x, y, width, height] in pixel coordinates with four integers. It should span the full image width and start at y=0, ending where the annotation strip begins. Record how you determined the boundary.
Illumination / vignetting correction. The micrograph has uneven illumination with darker corners and a measurable corner-versus-center bias. Correct it before any intensity-based thresholding. Describe the flat-field, background-estimation, or top-hat method and its parameters reproducibly (e.g., filter size, polynomial degree, rolling-ball radius). Explain why correction matters and quantify its effect using your own masks or region-wise statistics.
Pore segmentation and artifact rejection. Segment true metallurgical pores (dark rounded/irregular pores throughout the field) and exclude preparation contamination: a long polishing scratch and isolated surface debris specks. Save a final binary pore mask at original image dimensions. Explain the objective rule used to distinguish defects from pores (shape, size, linearity, location, manual curation) and any size or eccentricity filters. Pore, boundary, and artifact intensities overlap, so do not treat a single raw global threshold as a complete segmentation method.
Grain separation and spatial labeling. Segment individual grains over the full analysis ROI and assign every ROI pixel to exactly one grain territory, including pore and boundary pixels. The field includes locally weak boundaries and touching grains (notably in the upper-right). Use marker-controlled watershed or an equivalent marker-based separation to split touching grains while not over-splitting; a component count alone is not a sufficient result. Include edge grains clipped by the analysis ROI consistently and state the policy explicitly. Flag per-grain edge contact in the CSV. Produce a 16-bit label image where each positive integer maps to one grain.
Measurements. For each final grain defined by the label image, measure area in pixels and in um2, centroid in native image coordinates, and equivalent-circle diameter equivalent_diameter_um = sqrt(4 * area_um2 / pi). Produce one row per grain. Compute porosity as:
porosity_percent = 100 * (pore_pixels_inside_ROI) / (total_pixels_inside_ROI)
where total_pixels_inside_ROI includes pore pixels (pores remain in denominator). Do not include annotation-strip pixels or the scale bar in this denominator. Summarize distribution.
Planimetric-style estimate. Report grain_count = number of positive IDs in the label image = number of CSV data rows, analysis_area_um2 = ROI pixel area * calibration^2, and grain_density_per_mm2 = grain_count / (analysis_area_um2 / 1e6). Convert that density to an approximate ASTM E112 grain-size number using G = 1 + log2(grain_density_per_mm2 / 15.5). Label both as single-field planimetric-style estimates, not formal ASTM E112 compliance. Also report mean and median ECD in um.
Create artifacts/ and save exactly these seven files there:
measurements.csv: Exactly one row per segmented grain after final separation, including edge grains, and exactly one row per positive ID in grain_labels.png. Columns in exact order:
grain_id,centroid_x_px,centroid_y_px,area_px,area_um2,equivalent_diameter_um,touches_edge
grain_id unique integer 1..N contiguous, centroid_x_px/centroid_y_px numeric pixel coordinates computed from final label pixels in native image coordinates, area_px integer count of pixels with that ID in grain_labels.png, area_um2 numeric, equivalent_diameter_um numeric, touches_edge string true or false lowercase indicating whether that grain's label pixels touch the analysis ROI border. No extra columns. Each grain_id must equal its pixel value in grain_labels.png; area_px is the count of those pixels; centroid and edge flag derive from those pixels.
summary.json: JSON object with numeric fields:
calibration_um_per_px, analysis_roi_xywh_px (array of four integers), analysis_area_um2, grain_count, grain_density_per_mm2, astm_e112_grain_size_number_estimate, porosity_percent, mean_equivalent_diameter_um, median_equivalent_diameter_um
plus string fields:
edge_grain_policy (describe inclusion), porosity_denominator (must state that denominator includes pores and excludes annotation strip/scale bar).
Use finite JSON numbers, not strings or NaN/Inf; fields defined as integers must remain integers. Serialize units as ASCII um and um2 in keys only as given — do not use µm. grain_count must equal number of positive IDs in grain_labels.png and row count of measurements.csv.
pore_mask.png: 8-bit grayscale binary image at original image dimensions (same width/height as input micrograph, 1024x768). Pore pixels = 255, all other pixels = 0, including the entire bottom annotation strip and background. No intermediate gray values, anti-aliasing, or alpha channel. Every pixel where grain_labels.png belongs to a grain still contributes to the denominator; pore mask and label image are defined over the same ROI but stored separately.
grain_labels.png: Spatial grain territory map at original image dimensions (1024x768). Requirements: unsigned 16-bit grayscale PNG (mode I;16 or equivalent 16-bit), zero throughout the annotation strip (y>=704), and positive integer IDs 1..N contiguous in the full 1024x704 ROI. Every ROI pixel must belong to exactly one grain territory, including pore and boundary pixels; no zero/unassigned pixels are allowed inside ROI. IDs are arbitrary but must be contiguous 1..N and consistent with measurements.csv (each CSV grain_id equals its pixel value; area and centroid derived from those pixels). Grain count equals number of positive IDs.
segmentation_overlay.png: The original micrograph with grain boundaries and pore regions clearly distinguished in contrasting colors (e.g., cyan boundaries, red pores). Show the excluded polishing scratch / debris in a third color or explicitly annotate it. Preserve the source image content and the visible scale bar label/bar. The image must be human-auditable and consistent with both pore_mask.png and grain_labels.png.
grain_size_histogram.png: Publication-quality histogram of equivalent_diameter_um. Requirements: labeled x-axis "Equivalent diameter (um)" and y-axis "Count" or frequency, units in label, visible mean and median markers with numeric values or legend, and sample count (grain_count) annotated on figure or caption.
report.md: Concise engineering report covering:
um_per_px;xywh and how annotation-strip boundary found;G (both labeled single-field planimetric-style estimates), porosity, mean/median ECD;porosity_percent = pore_ROI / total_ROI * 100.pore_mask.png and grain_labels.png.touches_edge.measurements.csv contains exactly one row per final grain and per positive ID in grain_labels.png, grain_count == row count == number of positive IDs.grain_id equals its pixel value in grain_labels.png; area_px is the count of those pixels; centroid and edge flag derive from those label pixels in native coordinates.grain_labels.png is original image size, unsigned 16-bit, contiguous IDs 1..N, every ROI pixel assigned exactly one grain, including pore and boundary pixels; no zero/unassigned inside ROI.pore_mask.png is original image size, binary 0/255 only.um and um2.um / um2.measurements.csv centroid_x_px/centroid_y_px/area_px/area_um2/equivalent_diameter_um must be internally consistent with grain_labels.png label counts and centroids.artifacts/ and are readable, then stop.Images load automatically. YAML loads when expanded; audio and other formats remain metadata-only.
8 evaluation criteria
Porosity and pore mask fidelity including vignetting and artifact rejection [TOP PRIORITY - outranks grain count]: Compare pore_mask.png to for_graders/true_pore_mask.png (13053 true pore px, truth porosity 1.810663% in answer_key.json) and segmentation_overlay.png to for_graders/reference_overlay.png where red=true pores, yellow=planted scratch/debris, cyan=boundaries. Requirements: pore_mask.png same HxW as input (1024x768), 8-bit grayscale binary only 0/255 with no alpha, zero everywhere for y>=704; recomputed porosity must match summary.json. The truth band is 1.610663-2.010663%, narrower than the 0.279513 pp artifact-only area. Must reject the long scratch and two specks, and vignetted corners must not become false pores. Revised pixels overlap class intensities: the best raw threshold has Dice 0.813 and the best raw threshold inside the porosity band has Dice 0.810. Strong work therefore requires pore-mask Dice >=0.88 plus in-band porosity and artifact exclusion; Dice 0.82-0.88 is partial, and <=0.82 does not demonstrate the intended workflow. report.md must document reproducible illumination correction and evidence-based attribution. No naive percentage is a grading target. If A reaches the strong mask band with artifacts excluded and B merely lands on porosity with a raw-like mask, prefer A.
Spatial grain-label fidelity and segmentation overlay auditability [HIGH - label fidelity outranks exact count]: grain_labels.png must be original 1024x768 unsigned 16-bit grayscale, zero for y>=704, contiguous positive IDs 1..N covering every ROI pixel including pores and boundaries. CSV IDs, pixel areas, centroids, and edge flags must derive from this map. Compare submitted labels to for_graders/true_grain_labels.png after optimal one-to-one overlap matching; IDs need not equal truth IDs. Use for_graders/true_grains.csv for centroid and area checks. Strong work should achieve area-weighted matched IoU >=0.70 and boundary F1 >=0.75 with a 2 px tolerance; lower values receive proportional partial credit. Penalize merged super-grains, severe splits, or missing ROI coverage even when count is in band. The overlay must preserve the source and scale bar and distinguish boundaries, pores, and excluded artifacts across the full ROI including the weak upper-right zone. If A has higher matched spatial fidelity and B only has a closer count, prefer A.
Grain count and planimetric density / ASTM G / ECD distribution [strict +/-8 but lower priority than pore-mask quality and spatial label fidelity]: Ground truth true_grain_count=63 in answer_key.json from seeded Voronoi. Tolerance: grain_count_absolute_tolerance=8 => 55-71 acceptable inclusive, supported by controlled flat-field check supporting 55-66 territories; spatial label fidelity remains required because count alone is not sufficient. Edge grains clipped by ROI border must be included, not dropped, and flagged via touches_edge in CSV. grain_count must equal len(measurements.csv) and number of positive IDs in grain_labels.png. Truth grain_density_per_mm2=349.564986, astm_e112_G=5.495221. grain_density must equal grain_count/(analysis_area_um2/1e6), G must equal 1+log2(density/15.5) within rounding and within astm_e112_G_absolute_tolerance=0.5 (4.995-5.995 approx). Mean/median ECD truth 60.049919/60.027226 um, tolerance ecd_mean_relative_tolerance=0.15 and ecd_median_relative_tolerance=0.15 => ~51.0-69.1 um when calibration correct. If A has pore-mask within 0.20 pp and spatial labels good with count 60 within band and B has count 63 exact but raw-component-level territories or includes yellow scratch as pores failing porosity, prefer A because pore-mask and label fidelity outrank exact count. Explicitly state that count inside +/-8 is insufficient if grain_labels.png has merged/split/misregistered territories.
Per-grain schema and distribution accuracy + histogram auditability: measurements.csv must have exact header grain_id,centroid_x_px,centroid_y_px,area_px,area_um2,equivalent_diameter_um,touches_edge in order, one row per final grain and per positive ID in grain_labels.png, grain_id unique 1..N contiguous, touches_edge exactly 'true'/'false' lowercase. centroid_x_px/centroid_y_px numeric from grain_labels.png pixels, area_px equals count of those pixels, area_um2 = area_px * calib^2, equivalent_diameter_um = sqrt(4*area_um2/pi) per row within rounding. Consistency check: recompute centroid and area from grain_labels.png and match CSV within 1 px for centroid and exact for area_px. Mean/median ECD recomputed from CSV must be within 15% of truth when calibration correct. grain_size_histogram.png must have x-axis labeled 'Equivalent diameter (um)' or 'um', y-axis 'Count' or frequency, units in label, visible mean and median markers with numeric values or legend, and sample count (grain_count) annotated. If A has correct header/schema and distribution within tolerance and consistent with labels and B has swapped columns, extra columns, or mean/median far off despite calibration correct, prefer A.
Summary.json schema, scientific contracts, and ASTM labeling: summary.json must contain numeric fields calibration_um_per_px, analysis_roi_xywh_px (array of 4 integers), analysis_area_um2, grain_count, grain_density_per_mm2, astm_e112_grain_size_number_estimate, porosity_percent, mean_equivalent_diameter_um, median_equivalent_diameter_um and string fields edge_grain_policy (must state inclusion of edge grains) and porosity_denominator (must state pores remain included in denominator and annotation strip/scale bar excluded). Values finite JSON numbers not strings or NaN/Inf; analysis_roi equals that used for area and porosity. Grain density and G labeled as single-field planimetric-style estimates, not formal ASTM compliance. Missing fields or alternate porosity denominator makes output substantially worse. grain_count must equal CSV rows and positive IDs in grain_labels.png. If A meets schema and contracts and B missing fields or wrong denominator, prefer A.
Methodology and evidence-based naive/raw QC with attribution [reproducibility]: report.md must contain measured scale-bar length in px and derived um_per_px, exact ROI xywh and how annotation boundary found, illumination/background correction method with concrete parameters and quantified corner-vs-center improvement, pore cleanup/artifact-exclusion logic (size/shape/linearity/manual curation for scratch/debris), grain separation and spatial labeling method (marker-controlled watershed or equivalent with marker generation and full ROI assignment including pores/boundaries), edge-grain policy, results (calibration, ROI, area, grain count with label fidelity discussion, density and approximate ASTM E112 G labeled single-field planimetric-style estimates, porosity, mean/median ECD), limitations. Must report numeric naive uncorrected global-threshold comparison: at least naive porosity and naive grain count or bias versus final, with evidence-based attribution separating observed contributions from vignetting, boundaries, scratch/debris, and weak-boundary merging using evidence from its own masks or region-wise statistics. Correct attribution may conclude a factor is negligible for chosen threshold but must show measurement supporting claim. Do NOT prescribe or reward any expected naive percentage especially ~6.8% old dark fraction (previously boundaries/halo/artifacts with 0.0000 pp vignetting at dead band). Prefer concrete parameters and quantified bias over generic claims; do not reward claims of formal ASTM compliance.
Cross-file numerical consistency and binary/label encoding [exact]: grain_count == len(measurements.csv) rows == number of positive IDs in grain_labels.png; centroid_x_px/centroid_y_px recomputed from grain_labels.png within 1 px of CSV; area_px recomputed from grain_labels.png equals CSV exactly; area_um2 recomputed from area_px and calibration within 1%; mean/median ECD recomputed from CSV must match summary.json within 0.5 um after rounding and match histogram annotations within 1 um; analysis_area_um2 equals ROI widthheightcalib^2 within 1%; porosity_percent equals 100*(pore_mask 255 pixels inside ROI)/ROI pixel count within 0.2 pp; pore_mask original dimensions 1024x768 verified, binary only 0/255 (allow tol 1 for compression but prefer exact), zero in annotation strip; grain_labels.png original dimensions 1024x768, unsigned 16-bit, zero in annotation strip, contiguous IDs 1..N, every ROI pixel assigned exactly one ID, no zero inside ROI; histogram sample count matches grain_count. Internal contradictions (CSV 63 but labels 62, or porosity 3% but mask counts to 1.9%) make output substantially worse. Prefer exact encoding.
Deliverable completeness and file validity + visual auditability final: All seven required files must exist under artifacts/ exactly named measurements.csv, summary.json, pore_mask.png, grain_labels.png, segmentation_overlay.png, grain_size_histogram.png, report.md and readable non-zero, PNGs Pillow-openable, CSV valid with header grain_id,centroid_x_px,centroid_y_px,area_px,area_um2,equivalent_diameter_um,touches_edge, JSON valid finite numbers. segmentation_overlay.png must preserve original content and scale bar label and use distinguishable colors; grain_size_histogram.png publication-quality labeled axes with um units, readable ticks, mean/median markers, sample count. grain_labels.png must be 16-bit with contiguous IDs and full ROI coverage. Missing blank corrupt artifact major failure; if A has all seven valid and B missing pore_mask, grain_labels.png, or measurements.csv, prefer A even if B report longer. Do not penalize extra files, only require seven. Scientific correctness outranks visual polish, but unreadable figures fail auditability.