1. Start from the release, not from “latest”
For reproducible experiments, record the exact RoofVIP version you use. The Dataset page lists released and provisional versions separately.
2. Choose the sample organization you need
The original v1.0.0.0 benchmark is primarily object-centred. v1.0.3.0 instead provides fixed 256 × 256 pixel scenes, retaining multiple buildings and partial buildings where they intersect scene boundaries.
3. Choose modalities explicitly
For RGB + roof-vector experiments, the original release may be sufficient. For elevation-, point-cloud-, or multimodal experiments, v1.0.3.0 provides co-registered RGB, DSM, DTM, nDSM, LiDAR, and vector products using common scene identifiers.
4. Use the predefined split for v1.0.3.0 benchmarking
Version 1.0.3.0 provides 8,195 training, 1,022 validation, and 1,028 test scenes. The geographic domain is partitioned before tiling to reduce spatial leakage, and the subsets are balanced using PCA-based structural-complexity descriptors.
5. Record the benchmark protocol separately
When using the RoofVIP Benchmark evaluation layer, report both the dataset release and the benchmark protocol version. The dataset release v1.0.3.0 is available now; the formal evaluation protocol and leaderboard specification remain under development.