Bridging Annotation Granularity in Synthetic-to-Real Forest Tree Segmentation: The MGTD Dataset

Forest inventory is an essential task for Carbon stock estimation and sustainable woodland management. Yet, ground-based tree instance segmentation still remains bottle-necked by the prohibitive cost of fine-grained labeling. To cope up with this, we introduce MGTD, a mixed-granularity dataset of 53k synthetic forest images (with fine-grained annotations i.e. tree trunk and whole tree) along-with 3.6k real images (with coarse label i.e. tree). This exposes 2-coupled challenges; sim-to-real domain shift and annotation-granularity mismatch. To study them systematically, we propose a 4-stage evaluation protocol and a granularity-aware distillation baseline that transfers structural priors from fine-grained synthetic teachers to a student trained on coarse real labels. On real-world forest images, distillation improves mask AP from 0.396 to 0.429 over the same ResNet-50 real-only baseline. These results further provide a practical route towards more scalable ground-based forest perception.