Identify useful dataset parameters for semantic segmentation pre-training

Determine which dataset parameters are useful for pre-training semantic segmentation networks, including factors such as the number of classes and the presence of occlusions, in order to guide efficient construction of synthetic pre-training datasets and improve downstream segmentation performance.

Background

Semantic segmentation pre-training is valuable because dense pixel-level labeling is labor-intensive, making large annotated datasets costly to produce. Synthetic and formula-driven datasets can avoid manual annotation, but their effectiveness depends on selecting appropriate dataset parameters.

The authors note that it is not yet established which parameters (e.g., number of classes, occlusion) are most beneficial for pre-training segmentation models. They introduce SegRCDB and a framework to empirically investigate seven controllable factors, aiming to clarify which parameters matter and how they affect performance.