We present the Mok­ume data­set for sol­id wood tex­tur­ing con­sist­ing of 190 cube-shaped samples of vari­ous hard and soft­wood spe­cies doc­u­mented by high-res­ol­u­tion ex­ter­i­or pho­to­graphs, an­nu­al ring an­nota­tions, and volu­met­ric com­puted tomo­graphy (CT) scans. A sub­set of samples fur­ther in­cludes pho­to­graphs along slanted cuts through the cube for val­id­a­tion pur­poses.

Us­ing this data­set, we pro­pose a three-stage in­verse mod­el­ing pipeline to in­fer sol­id wood tex­tures us­ing only ex­ter­i­or pho­to­graphs. Our meth­od be­gins by eval­u­at­ing a neur­al mod­el to loc­al­ize year rings on the cube face pho­to­graphs. We then ex­tend these ex­ter­i­or 2D ob­ser­va­tions in­to a glob­ally con­sist­ent 3D rep­res­ent­a­tion by op­tim­iz­ing a pro­ced­ur­al growth field us­ing a nov­el iso-con­tour loss. Fi­nally, we syn­thes­ize a de­tailed volu­met­ric col­or tex­ture from the growth field. For this last step, we pro­pose two meth­ods with dif­fer­ent ef­fi­ciency and qual­ity char­ac­ter­ist­ics: a fast in­verse pro­ced­ur­al tex­ture meth­od, and a neur­al cel­lu­lar auto­maton (NCA). We demon­strate the syn­ergy between the Mok­ume data­set and the pro­posed al­gorithms through com­pre­hens­ive com­par­is­ons with un­seen cap­tured data. We also present ex­per­i­ments demon­strat­ing the ef­fi­ciency of our pipeline's com­pon­ents against ab­la­tions and baselines. Our code, the data­set, and re­con­struc­tions are avail­able via ht­tps://mok­umepro­ject.git­hub.io/