ditr-industrial-23cls

DITR (PTv3 + frozen DINOv2-small feature injection) semantic segmentation model for industrial scenes, trained on synthetic Isaac Sim point cloud data with 23 classes (floor, wall, rack, pallet, forklift, robot, ...).

Test result: mIoU 0.8563 / mAcc 0.8796 / allAcc 0.9879.

Contents

• ditr-industrial-23cls.pth: slimmed weights ({"state_dict": ...}, DINOv2 included)
• config.py: merged training config (model, class names, data pipeline)
• industrial_rgbd.py: custom dataset class required by the config
• md5sums.txt: file checksums

Dataset and detailed reproduction instructions are not published yet; documentation will be extended later.