Computer vision.
Built for production.

My focus is building a robust object detector within Scene Parsing. The challenge spans identifying individual products in shelf imagery and supporting product recognition within the broader ShelfView system.

I work with product and data teams to connect model development and evaluation to the application’s needs.

The pipeline used a custom 2D U-Net for the initial knot segmentation and a custom 3D U-Net with volumetric convolutions for refinement. Downstream algorithms used the resulting masks to infer parametric representations of individual knots.

Training examples came from multiple customers, connecting model development to the variety of wood encountered in industrial scanning. My broader workflow covered data preparation, augmentation, architecture selection, and deployment, balancing precision with production-time constraints.

For lumber, the work included knot and stain segmentation on dry boards and lengthwise grain classification. I also explored multi-defect segmentation on longitudinal scanners using RGB, laser-scatter, 3D, and X-ray inputs. These multi-defect experiments were exploratory at the time of the presentation.

For dates, the tasks included quality sorting, ripeness, and fungus classification using RGB and near-infrared imagery. Apple inspection distinguished good stems, cracked stems, and calyx regions. This connected visual recognition to practical quality decisions, with data collected and labeled alongside domain specialists.