ditr-industrial-aligned-23cls

DiTR Industrial Aligned 23-Class

DiTR semantic-segmentation checkpoint combining PTv3 geometry features with a frozen DINOv2-small RGB encoder. It was trained on aligned synthetic Isaac Sim industrial RGB-D data.

The checkpoint metadata records best_metric_value=0.8728377743924077 at epoch 58. This value comes from the training experiment's configured validation metric; reproduce evaluation on the intended dataset before using it as a benchmark.

Files

• ditr-industrial-aligned-23cls.pth: inference-only task state_dict
• config.py: merged Pointcept/DiTR experiment configuration
• sha256sums.txt: integrity hashes

Optimizer, scheduler, and training history were removed. Frozen DINOv2-small weights are not duplicated in this task checkpoint; the deployment image fetches and caches them during the Docker build.

Classes

The fixed model order is:

barcode, bracket, cardboard_box, cnc_machine, container, conveyor,
fire_extinguisher, floor, floor_decal, forklift, lamp, pallet,
pallet_trolley, pillar, rack, robot, robot_stand, safety_fence, shelf,
sign, table, wall, workpiece

Reproducible API deployment

The Dockerized FastAPI server, RTX 4070/4080 settings, request contract, operations guide, and checksum verification are documented on the ditr branch:

• https://github.com/mingqian0850/industrial_3d_sem_seg_server/tree/ditr

The deployment pins DiTR commit f66d1dadb82e97ded7750567a043e4783305b01c.

git clone https://github.com/mingqian0850/industrial_3d_sem_seg_server.git
cd industrial_3d_sem_seg_server
git switch ditr
./deploy_scripts/build_api.sh
GPU_DEVICE=0 MAX_VALID_POINTS=350000 ./deploy_scripts/start_api.sh

Intended use and limitations

This model is intended for research and evaluation on RGB-D scenes compatible with the synthetic industrial training domain, camera projection assumptions, and published 23-class taxonomy. It may not generalize to unseen sensors, lighting, geometry, or class definitions. Review predictions before any safety-relevant use.