Instance Segmentation for Industrial Forestry Computer Vision
Challenge
The client builds computer vision systems for sawmills — automating log scanning, board grading, and quality control. They needed instance segmentation of individual logs in cross-section views, boards in stacked configurations, and sawmill scenes with overlapping objects. Distinguishing between foreground and background logs, handling overlapping boards, and classifying surface types required highly specialized annotation.
Solution
We built deep domain expertise in lumber-specific annotation across 10+ project types — from log face segmentation to board end-surface analysis. Close collaboration with the client's ML engineers ensured annotation standards evolved with their model requirements. We proactively analyzed annotation data and identified that removing optional label types could save 30% of annotation time — helping the client optimize their ML training budget. Typical turnaround of 2-7 days per batch.
Results
Annotation density averages ~280 polygon points per image — significantly higher than typical object detection tasks. We proactively analyzed annotation data and identified a 30% time-saving optimization by removing optional label types the client's models didn't need for deployment.