Moltline Vision Maths MCP Server

Convert bounding boxes between COCO, Pascal VOC and YOLO. FREE. The three formats disagree on everything: COCO is [x, y, width, height], VOC is [x1, y1, x2, y2], YOLO is [cx, cy, w, h] normalised to the image. Getting this wrong produces boxes that look plausible and quietly ruin every metric. Typical input {"boxes": [[10, 20, 100, 50]], "from_format": "coco", "to_format": "yolo", "image_width": 640, "image_height": 480} returns {"boxes": [[0.0938, 0.0938, 0.1562, 0.1042]], "converted": 1, "rejected": []}. Use whenever a dataset and a model disagree about format. Not for scoring predictions (detection_metrics) and not for removing overlaps (nms). Errors: on invalid, missing, or malformed input this tool never raises a protocol error — it returns {"error": "<what is wrong and how to fix it>"} (for example {"error": "boxes must contain at least one box"}). Every call is read-only and idempotent, so after correcting the input it is always safe to retry.

Check a colour pair against the WCAG contrast thresholds. FREE. Uses the WCAG 2 relative-luminance formula, so the number matches what an accessibility audit will report. Typical input {"foreground": "#767676", "background": "#ffffff"} returns {"contrast_ratio": 4.54, "AA": true, "AAA": false, "required": {"AA": 4.5, "AAA": 7.0}, "verdict": "Passes AA for normal text, fails AAA."}. Use when generating or auditing an interface. Not for converting colours between spaces and not for palettes. Errors: on invalid, missing, or malformed input this tool never raises a protocol error — it returns {"error": "<what is wrong and how to fix it>"} (for example {"error": "foreground must be a hex colour like #767676"}). Every call is read-only and idempotent, so after correcting the input it is always safe to retry.

Score detections against ground truth and show the working. PREMIUM (license). Greedy matching at the IoU threshold, highest-confidence prediction first, each ground-truth box matched at most once - the standard protocol. Reports per-class precision, recall and F1, and average precision by the all-points interpolation used by Pascal VOC 2010 onward. Typical input {"predictions": [{"box": [0,0,10,10], "label": "cat", "score": 0.9}], "ground_truth": [{"box": [1,1,11,11], "label": "cat"}]} returns {"overall": {"tp": 1, "fp": 0, "fn": 0, "precision": 1.0, "recall": 1.0, "f1": 1.0}, "per_class": {...}, "mAP": 1.0}. Use to compare two models on the same held-out set. Not for cleaning up a single model's overlapping output first - run nms before this. Errors: on invalid, missing, or malformed input this tool never raises a protocol error — it returns {"error": "<what is wrong and how to fix it>"} (for example {"error": "ground_truth must contain at least one box"}). Every call is read-only and idempotent, so after correcting the input it is always safe to retry.

Read an image's format and pixel size from its header alone. FREE. Dimensions live in the first few dozen bytes of PNG, JPEG, GIF, BMP and WebP, so a base64 prefix is enough - you do not need to send the whole file, and nothing is decoded. Typical input {"data_base64": "iVBORw0KG..."} returns {"format": "png", "width": 1920, "height": 1080, "aspect_ratio": 1.7778, "aspect_label": "16:9", "megapixels": 2.07, "orientation": "landscape", "bytes_inspected": 512}. Use to find out what you are dealing with before planning a resize. Not for pixel content - nothing here reads pixels - and not for EXIF. Errors: on invalid, missing, or malformed input this tool never raises a protocol error — it returns {"error": "<what is wrong and how to fix it>"} (for example {"error": "data_base64 must not be empty"}). Every call is read-only and idempotent, so after correcting the input it is always safe to retry.

Remove duplicate detections of the same object. PREMIUM (license). Greedy non-maximum suppression: keep the highest-scoring box, drop everything overlapping it above the threshold, repeat. Ties break on the earlier index, so the result is deterministic rather than dependent on sort stability. Typical input {"boxes": [[0,0,10,10],[1,1,11,11],[50,50,60,60]], "scores": [0.9, 0.8, 0.7]} returns {"keep": [0, 2], "suppressed": [{"index": 1, "by": 0, "iou": 0.6807}], "kept": 2}. Use after a detector that emits overlapping boxes. Not for scoring against ground truth (detection_metrics) and not for format changes (bbox_convert). Errors: on invalid, missing, or malformed input this tool never raises a protocol error — it returns {"error": "<what is wrong and how to fix it>"} (for example {"error": "boxes and scores must be the same length"}). Every call is read-only and idempotent, so after correcting the input it is always safe to retry.

Work out the exact scale, padding and crop for a model input size. FREE. Returns the numbers you need to transform boxes alongside the image, which is the step that usually gets skipped. Typical input {"width": 1920, "height": 1080, "target": "yolo_640"} returns {"scale": 0.3333, "resized": [640, 360], "pad": {"left": 0, "top": 140, "right": 0, "bottom": 140}, "box_transform": "x_new = x * 0.3333 + 0; y_new = y * 0.3333 + 140"}. Use before feeding an image to a fixed-input model. Not for finding out the image's size in the first place - that is image_probe. Errors: on invalid, missing, or malformed input this tool never raises a protocol error — it returns {"error": "<what is wrong and how to fix it>"} (for example {"error": "unknown target <value>; use one of <value> or set"}). Every call is read-only and idempotent, so after correcting the input it is always safe to retry.