MIRA — Day 15: Repository Cleanup, EXP-019, and Reproducible Results

This time Instead of adding more features, I concentrated on making the existing software easier to understand, easier to reproduce, and just better overall. The main outcomes were a big repo cleanup, New kaggle results with EXP018/019 reaching all time high of mAP50 of 90%

Repository Cleanup and Reliability Improvements

I performed another detailed check of the complete codebase. Which highlightes some more issues as a lot of code was Ai generated. The review uncovered problems in both the runtime software and the research CLI/pipeline.

I cleaned up third party model files, duplicate documentation, archived ZIP files, and unused assets such as Datasets. The local repo had previously grown to ca 13 GB, largely because of old Git objects. After the cleanup, it was reduced to approximately 520 MB.

The README was reduced from roughly ** 900 lines** to approximately 340 lines. It now concentrates on the current installation process, the active perception pipeline, the best available models, dataset origins, known limitations, and reproducibility. Jugend forscht material was moved into the jufo/ directory, while the related-work section was separated into its own document and was put in gitignore for various reasons.

EXP-018 and EXP-019

After rebuilding the dataset and correcting the training workflow, I trained YOLO11n on the cleaned and balanced dataset. EXP-018 achieved the following results:

Metric Result mAP50 90.6% mAP50-95 82.2% Glass mAP50 90.8% Metal mAP50 94.8% Paper mAP50 88.7% Plastic mAP50 81.1% Trash mAP50 97.5%

Compared with EXP-014, which achieved 60.7% mAP50 on the older dataset, this represents an improvement of 29.9 percentage points.

I repeated the training as EXP-019 to test whether the result was reproducible. EXP-019 again achieved 90.6% mAP50 on the same clean dataset. So EXP-019 is now the best Model and the final one

Model Exports

The EXP-019 was exported into multiple formats for later deployment tests:

Format Approximate size Intended use FP32 PyTorch 5.5 MB Development and reference inference INT8 TFLite 2.90 MB Resource-constrained edge hardware ONNX 10.1 MB Cross-platform inference and benchmarking

The INT8 TFLite version is especially important for the planned Raspberry Pi Zero 2 W deployment. Its smaller size should reduce storage and memory pressure, although its real speed, temperature, and accuracy still need to be measured on the target hardware.

Next Steps

The next task is to benchmark EXP-019 on the Raspberry Pi Zero 2 W ordered through Stardance. The test will measure:

• Inference speed and end-to-end latency.
• Memory usage.
• Temperature during continuous operation.