Arman NeyestaniComputer Vision · MeasurementResume
← All work

Applied Vision

Blue crab segmentation and dimension estimation

Instance masks, aligned depth, and camera calibration are combined to estimate projected crab dimensions in millimeters.

Project repository ↗Dataset record ↗
My role
I developed this independent project, including dataset documentation, training configuration, inference code, examples, and a geometry-based measurement extension.
What I built
YOLO11m-seg predicts per-instance masks. Valid mask pixels are back-projected using depth and intrinsics, then PCA axes define two projected whole-mask extents.
Key results
  • 85.7% test mask mAP@50 (26-image re-evaluation)
  • 67.6% test mask mAP@50–95 (small held-out split)
Stack
Python · YOLO11m-seg · PyTorch · OpenCV · Depth imaging · PCA geometry
Female blue crab with an instance segmentation mask and confidence label.
Original repository example showing the visible animal boundary and predicted instance.

Problem

What needed to be understood or measured

Marine monitoring needs both animal identification and physical dimensions; segmentation alone provides outlines in pixels, not measurements.

My contribution

My contribution

I developed this independent project, including dataset documentation, training configuration, inference code, examples, and a geometry-based measurement extension.

Experimental setup

Data and setup

A 631-image, four-class dataset with separate train, validation, and test splits. The measurement path requires RGB-aligned metric depth and a calibrated intrinsic matrix.

Method

From input to interpretable output

YOLO11m-seg predicts per-instance masks. Valid mask pixels are back-projected using depth and intrinsics, then PCA axes define two projected whole-mask extents.

Workflow combining RGB instance masks, aligned depth and camera calibration for dimension estimation.Open figure —
Project diagram

Measurement workflow from segmentation and depth to projected dimensions.

Blue_crabs repository.

Results

Reported results

85.7%

test mask mAP@50

26-image re-evaluation

67.6%

test mask mAP@50–95

small held-out split

Validation & uncertainty

Reference and uncertainty checks

The saved model was re-evaluated on a 26-image test split in September 2026. The dimension demonstration itself has not been validated against physical reference measurements.

Limitations & failure modes

Operating limits and failure modes

Underwater blue crab detection sequence including missed and occluded instances.Open figure —
Repository output

Demo frames retain missed detections and occlusions as visible failure cases.

Blue_crabs repository.

Occluded limbs, changing pose, depth outliers, calibration error, refraction, sparse masks, and class imbalance affect both boundary quality and projected extents.

The physical accuracy of the dimension estimator is not established. It needs calibrated depth, underwater optical calibration, and comparison against reference measurements before measurement claims are justified.

Technical stack

Tools selected for the measurement chain

  • Python
  • YOLO11m-seg
  • PyTorch
  • OpenCV
  • Depth imaging
  • PCA geometry

Sources & project links

Papers and code

Public repository with dataset, checkpoint metadata, test re-evaluation, and explicit caveats around source-image licensing and metric validation.

Project repository ↗Dataset record ↗