Research record

What changed between EXP-001 and EXP-019.

I changed the model and the data several times. The data change made the larger difference, taking the project from a 61% classifier to a 90.58% mAP50 detector.

Reference modelEXP-019

mAP5090.58%

mAP50-9582.15%

Three changes shaped the project.

The experiment numbers matter because each group answered a different question. I started by learning the classification problem, then changed the task, and finally changed the data.

ClassificationEXP-001 to EXP-004

I started with a custom CNN.

I wanted a baseline I could understand before using a larger pretrained model. MobileNetV2 improved the classification result, but classification could not locate several objects in one frame.

EXP-001: 61.00%
EXP-002: 84.28%
EXP-003: 87.42%
EXP-004: 87.42%

DetectionEXP-005 to EXP-017

YOLO showed me the data problem.

I moved to YOLOv8n and then YOLO11n so the model could locate objects. The scores moved up and down across dataset combinations. Adding sources did not automatically make the detector more reliable.

EXP-005: 82.3%
EXP-006: 39.4%
EXP-014: 60.7%
EXP-017: 59.3%

Dataset changeEXP-018 and EXP-019

The smaller, cleaner set worked better.

I removed inconsistent examples and built a more balanced dataset. EXP-018 reached 90.6% mAP50, and EXP-019 repeated the result at 90.58%.

EXP-018: 90.60%
EXP-019: 90.58%

The data changed the result more than the model did.

EXP-018 and EXP-019 reached about 90.6% mAP50 after I removed inconsistent examples and rebuilt the detector around a cleaner dataset.

mAP50-95: 82.15%
Precision: 87.2%
Recall: 84.6%
Classes: 5

See the experiment record

Detection experiments

mAP50 progression

EXP-019 · 90.58%

Current through EXP-019. Values are generated from the repository experiment log.

The current reference model.

YOLO11n trained for 120 epochs on the clean balanced dataset, then exported to PyTorch, ONNX, and quantized TFLite.

Evaluation

mAP50: 90.58%
mAP50-95: 82.15%
Precision: 87.2%
Recall: 84.6%

Dataset

Train: 5,108 images
Validation: 415 images
Test: 1,375 images
Classes: 5 materials

Exports

PyTorch: 5.47 MB
ONNX: 10.61 MB
TFLite 320: 3.02 MB
TFLite 640: 3.04 MB

The current result still has limits.

Crumpled paper

White crumpled paper can be classified as plastic with high confidence.

Can orientation

Metal cans facing the camera opening-first can disappear from the detection.

Occlusion

Stacked objects reduce bounding-box quality, especially for paper and trash.

Follow the result back to the work.

Training pipelineRead the guide Source and testsInspect on GitHub