YCB Berkeley — YOLO26 Segmentation (Fine-tuned)

Instance segmentation model for YCB objects (YOLO26n-seg), fine-tuned on custom annotations with YOLO26 augmentations.

Fine-tuned YOLO26n-seg model for segmenting YCB objects in images or video. Trained on our own annotated data with YOLO26 augmentations, starting from a model pretrained on the YCB Berkeley dataset. Includes weights and validation figures; no dataset or training code included.

Model

• Weights: weights/best_finetuned_latest.pt
• Architecture: YOLO26n-seg (Ultralytics)
• Input size: 640×640
• Task: Instance segmentation

Classes (15)

bowl, cleanser, coffee_can, cracker_box, gelatin_box, jaco, meat_can, mug, mustard_bottle, rubik, soup_can, tennis_ball, windex, wooden_block, workspace

Usage

pip install ultralytics

from ultralytics import YOLO

model = YOLO("weights/best_finetuned_latest.pt")

# Image
results = model.predict("image.jpg")

# Webcam
results = model.predict(source=0, show=True, conf=0.25)

Results

Training and validation figures are in figures/:

• results.png — training curves
• confusion_matrix.png / confusion_matrix_normalized.png — classification confusion
• BoxF1_curve.png, BoxPR_curve.png — box metrics
• MaskF1_curve.png, MaskPR_curve.png — mask metrics
• val_batch0_pred.jpg, val_batch1_pred.jpg — sample validation predictions

License

This project is released under the MIT License. Weights and figures are provided for research and non-commercial use. Dataset and training code are not included in this repository.