Current ink labels are generated by a human annotator drawing the letter they think the ink model most closely resembles. This offers some benefits in that we may catch additional ink unseen by humans, but also increases the risk of training the model to focus on the underlying surface rather than the ink itself. We are looking for as-accurate-as-possible ink labels that are representative of only the detectable ink patterns on the surface, and in true 3d rather than a single image projected across multiple layers.
File Format: Zarr or .tif slices
Result: Image/label pairs of high quality 3d ink labels that are “ready-to-run” , no additional cropping or preprocessing required.
Current ink labels are generated by a human annotator drawing the letter they think the ink model most closely resembles. This offers some benefits in that we may catch additional ink unseen by humans, but also increases the risk of training the model to focus on the underlying surface rather than the ink itself. We are looking for as-accurate-as-possible ink labels that are representative of only the detectable ink patterns on the surface, and in true 3d rather than a single image projected across multiple layers.
File Format: Zarr or .tif slices
Result: Image/label pairs of high quality 3d ink labels that are “ready-to-run” , no additional cropping or preprocessing required.