SewerSim Full Dataset for Sim-to-Real Sewer-Tunnel Segmentation

Project Context: REISAR and France 2030 This dataset release is part of the REISAR project (Advanced Robotic System for Sewer Network Inspection and Water Preservation), funded under the ANR-23-DMRO-0014 reference and supported by the France 2030 investment plan. REISAR addresses robotic inspection and mapping challenges in sewer networks, which are difficult to access, visually degraded, and hazardous for human operators. Within this context, the present dataset supports research on sim-to-real perception for sewer-tunnel inspection through annotated real and synthetic image data for segmentation tasks. It accompanies the SewerSim study on synthetic data generation and sim-to-real instance segmentation in confined sewer tunnels. The publication archive is built from the complete SewerSim benchmark release used for the paper and reorganized for clean archival publication on Zenodo. The benchmark targets two semantic instance classes: debris tunnel and preserves the three evaluation views used in the study: real_only sim_only simreal Technical Contribution: Full Public Release of the SewerSim Benchmark This archive provides: full YOLO-format image and label datasets COCO-format instance-segmentation annotations split-level statistics and release metadata integrity manifests for upload and download verification Summary Table of Dataset Contents Variant Images Split structure Resolution Notes real_only 2,594 1,794 train / 300 val / 500 test 1280x720 real tunnel-run dataset sim_only 8,750 6,510 train / 1,037 val / 1,203 test 1920x1200 synthetic Isaac Sim dataset simreal 11,344 8,304 train / 300 val / 500 test, plus sim_val and sim_test mixed: 1920x1200 train, 1280x720 val/test mixed-domain training protocol Notes for Reuse simreal is a protocol composition rather than an independent new corpus configs/simreal.yaml uses only train, val, and test sim_val and sim_test are retained as synthetic reference splits COCO annotations are provided without duplicated image files