MIRA-AI — Machine Intelligence for Recycling Automation
Trained models for the MIRA waste sorting project (Jugend Forscht 2027). All models target 5 material classes: glass, metal, paper, plastic, trash.
Stage A — Image Classification (MobileNetV2)
| File |
Val Acc |
Size |
Notes |
mira_classifier_baseline.keras |
61.00% |
45.71 MB |
3-layer custom CNN baseline |
mira_classifier_transfer.keras |
84.28% |
9.25 MB |
MobileNetV2 frozen base |
mira_classifier_tuned.keras |
87.42% |
23.48 MB |
MobileNetV2 fine-tuned, best Keras accuracy |
mira_classifier_fp32.tflite |
87.42% |
8.49 MB |
FP32 TFLite export |
mira_classifier_int8.tflite |
87.42% |
2.61 MB |
INT8 quantized — best for edge deployment |
Stage B — Object Detection (YOLO)
| File |
mAP50 |
Params |
Size |
Training Data |
mira_exp014.pt |
60.7% |
2.58M |
5.21 MB |
TACO + TrashNet + Roboflow (6,802 img) — CURRENT BEST |
mira_exp014_int8.tflite |
60.7% |
2.58M |
2.90 MB |
INT8 quantized version |
mira_exp017.pt |
59.3% |
2.58M |
5.21 MB |
All 4 sources (TACO + TrashNet + Roboflow + WaRP, 9,774 img) |
mira_exp016.pt |
58.8% |
2.58M |
5.21 MB |
WaRP only |
mira_exp015.pt |
56.0% |
2.58M |
5.21 MB |
TACO + TrashNet + WaRP |
mira_exp013.pt |
55.1% |
2.58M |
5.21 MB |
TACO + TrashNet (4,024 img) |
mira_exp006.pt |
39.4% |
3.01M |
5.94 MB |
YOLOv8n, wild + TrashNet |
mira_exp011.pt |
35.0% |
3.01M |
5.94 MB |
YOLOv8n, TACO only |
mira_exp009_int8.tflite |
72.8% |
3.01M |
3.18 MB |
Tabletop only (inflated by clean backgrounds) |
Other Models
| File |
Description |
mira_exp006_int8.tflite |
INT8 quantized YOLOv8n |
mira_exp011_int8.tflite |
INT8 quantized TACO-only |
mira_exp013_int8.tflite |
INT8 quantized YOLO11n |
mira_exp015_int8.tflite |
INT8 quantized TN+WaRP |
mira_exp016_int8.tflite |
INT8 quantized WaRP-only |
mira_exp017_int8.tflite |
INT8 quantized all-4-sources |
mira_exp017.onnx |
ONNX export |
gianlucasposito_yolov8n.pt |
Third-party benchmark model |
Key Findings
- More data ≠ better — adding WaRP to the 3-source mix (EXP-017) dropped mAP50 from 60.7% to 59.3%
- Trash class is the bottleneck (26.9% mAP50) — extreme intra-class diversity
- INT8 quantization achieved 3.5x compression (10.14 MB → 2.90 MB) with no measurable accuracy loss
- All models target Raspberry Pi Zero 2W edge deployment
Source Code
github.com/jeremy341/MIRA-AI
