Astra Vision V1
Construction site safety & material detection — detects 8 classes: steel bar, cement bags, concrete blocks, red bricks, no helmet, no safety vest, helmet, safety vest.
| Class | AP |
|---|---|
| cement_bags | 77.2% |
| red bricks | 64.8% |
| helmet | 49.2% |
| safety_vest | 47.4% |
| steel bar | 39.4% |
| no_helmet | 39.4% |
| concrete_blocks | 38.7% |
| no_safety_vest | 21.8% |
mAP: 47.23% (AP@IoU=0.50:0.95) on 10,865 validation images
Architecture: YOLOX-S (9.90M params)
Input: 640x640
Usage
import torch
ckpt = torch.load("astra_vision_v1_best.pth", map_location="cpu")
model.load_state_dict(ckpt["model"])
Training
- 50 epochs, batch size 8, SGD optimizer
- CIoU loss, cosine LR schedule
- 51,970 train + 10,865 val images (1,580,827 annotations)
- NVIDIA RTX 3060 12GB
Links
- Source: https://github.com/Astramaan/astra-vision
- Weights: https://huggingface.co/Astramaan/astra-vision
- Demo: This Space
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support