- Roboflow dataset (Original) - link
- Dataset (Custom Augmented) - link
- Model training version 1 - link
- Version 1 with totally 110 data samples in yolo11s.
- Version 1 inference
- Annotate Manually
- Populate the dataset to 1000+
- Accuracy low (suspect: low data samples)
- Version 2 trial with 900 training samples (Manual Annotation with Augmentation [300x3] via roboflow) in yolo11m
- Version 2 inference
- Tested with 900 training samples (300 annotated x 3 r-augments) with much better decent results but yet to achieve the saturation!
- Shifting from (32,9,5) split to (96,9,5) split to finally (900,43,5) split has shown improved results. (train, valid, test)
- 'mAP50' has increased by around 82 times and 'mAP50-95' has increased by around 148 times.
- Version 3 trial with more data samples, yolo11l and 25e:100e
- Tested with 1374 training samples and split being (1374,50,0).
- yolo11x seemed to be too complex for this task and was easily overfitting - so neglected!
- yolo11l seemed comparatively better that yolo11x with decent results.
- Version 3 is the best possible saturation with yolo via cli. Trying pythonic way may (or may not) give better control over arguments that could pave way to enhanced results.
