min99ian/volt-industrial-23cls

Volt industrial 23-class semantic segmentation

This repository contains the deployment artifacts for a Volt-S model fine-tuned on a custom industrial RGB-D point-cloud dataset.

The matching Docker/FastAPI service is the volt branch of mingqian0850/industrial_3d_sem_seg_server. It implements the same isaac-capture.v1 request and response contract as the project's DiTR and PTv3 deployments.

Files

• config.py: frozen training/evaluation configuration.
• volt-industrial-23cls.pth: inference-only checkpoint containing the EMA state_dict.
• sha256sums.txt and md5sums.txt: artifact integrity checks.

The inference checkpoint was extracted from model_best.pth at epoch 98. Its recorded best validation mIoU is 0.9250934182924747. Optimizer, scheduler, scaler, and the non-EMA training weights were intentionally omitted from this deployment artifact.

Source details:

• Volt repository: https://github.com/mingqian0850/Volt.git
• Volt commit: 089cc38d8b32e7c695dd939057f787f3c60dd35e
• Original full-checkpoint SHA-256: 8793a07edd6777bf171f690e6f33fed7cdf3ad5d863819376ede4380ab4fa81f

Input and preprocessing

The model consumes XYZ coordinates plus six input features:

• RGB scaled from [0, 255] to [0, 1]
• surface normals (nx, ny, nz)

Deployment uses the frozen test pipeline from config.py: 2 cm grid voxelization, inverse mapping to the original valid depth points, XY centering, color normalization, and Volt tokenization with a 5-voxel patch size. The service uses a fixed voxel representative seed for repeatable requests.

Classes

Predictions are zero-based IDs in this order:

1. barcode
2. bracket
3. cardboard_box
4. cnc_machine
5. container
6. conveyor
7. fire_extinguisher
8. floor
9. floor_decal
10. forklift
11. lamp
12. pallet
13. pallet_trolley
14. pillar
15. rack
16. robot
17. robot_stand
18. safety_fence
19. shelf
20. sign
21. table
22. wall
23. workpiece

Integrity

6d5385d38c5d6060eacd3d226124c86804b8cee7e129664b4c365f3bd6e9ed36  config.py
65ba22f299847f2c4356a59f1f110160f865c7e9d82816d9e980161fe515c198  volt-industrial-23cls.pth

Intended use

This model is intended for the industrial scene domain represented by its training data. Performance under new sensors, layouts, materials, noise profiles, or class taxonomies should be validated before production use.