SimForest: RGBD Instance Segmentation Dataset

SimForest is a synthetic dataset generated using the AGRARSENSE forestry simulation environment developed in Unreal Engine 5. The dataset is formatted in COCO-style and includes aligned RGB and depth images, along with instance segmentation annotations. It is designed for training and evaluating models in scene understanding, semantic segmentation, RGB-D perception, and multimodal fusion, particularly in complex, natural forest environments. The dataset contains 5,000 high-resolution images at 3840×2160 pixels, captured with a virtual camera configured with a 90° field of view. The data covers all four seasons and a variety of cloudiness levels and daytime lighting conditions. Sun positions were dynamically simulated to correspond to the real-world solar angles of a geo-located forest area near Umeå, Sweden, adding realism and temporal diversity for tasks involving outdoor perception and environment modeling. SimForest includes 11 annotated classes, carefully selected to represent key elements in Nordic forest environments: Terrain – Ground surface such as soil, forest floor, and trails Foliage – Leafy vegetation and bushes. Birch, Pine, Spruce – Trees (including trunk, branches, and leaves), segmented by species Birch_Trunk, Pine_Trunk, Spruce_Trunk – Tree trunks, separated by species Sky – Sky and atmospheric background Rock – Static natural obstacles such as boulders, stones, and rocky terrain features Snow – Snow-covered terrain surface in winter scenarios In addition to visual and segmentation data, the dataset includes comprehensive metadata to support advanced perception and simulation research: Camera Metadata: Intrinsic parameters and full 6-DoF pose Object Metadata: Per-instance location, orientation, and physical dimensions Terrain Depth: A terrain depth map for each image, offering elevation information relative to the camera frame Weather Conditions: Season (Winter, Spring, Summer, Autumn) Time of day (hour) Month (affecting sun angle) Cloudiness, represented as a floating-point value between 0 (completely clear) and 1 (fully overcast) This dataset is particularly valuable for researchers working on sensor fusion, context-aware perception, domain adaptation, seasonal variation handling, and sim-to-real transfer learning in forestry and outdoor robotics.  Dataset extraction The dataset is compressed using 7-Zip and split into multiple volumes due to size limitations. A separate archive, SimForest_Metadata.7z, contains only the metadata. To extract the dataset download all SimForest.7z.00x files and place them in the same directory. Extract/open SimForest.7z.001. This will automatically join and unpack all volumes.  Supplementary material Supplementary scripts for converting the SimForest dataset to YOLO format, and for training and validation of YOLO detection and segmentation models are available in a public GitHub repository: https://github.com/RISE-Dependable-Transport-Systems/SimForest-YOLO-Toolkit