atod-acdc-b0-yolov8s

B0 YOLOv8s Original - ACDC

This repository contains the B0 historical baseline used in the ATOD-Generalization experiments.

Model

• Experiment: B0
• Architecture: YOLOv8s
• Initialization: pretrained weights
• Task: object detection
• Framework: Ultralytics 8.4.160
• Dataset: ACDC
• Training images: 1,600
• Validation images: 406
• Training instances: 10,135
• Validation instances: 2,794
• Number of classes: 8

Classes

Training Configuration

The same training recipe is used for the B0-B3 baseline comparison.

The official ACDC train and validation distributions are preserved without class rebalancing.

Official Validation Protocol

The reported metrics are obtained by re-evaluating the final best.pt checkpoint on the ACDC validation split after training.

This post-training validation protocol is used consistently for all B0-B3 baselines.

Overall Validation Results

Per-class Results

Training Metrics

Per-class mAP50-95

Validation Curves

Precision-Recall Curve

F1 Curve

Precision Curve

Recall Curve

Confusion Matrix

Normalized Confusion Matrix

Files

• best.pt: best validation checkpoint
• baseline_metrics.json: official post-training validation metrics
• experiment_snapshot.json: experiment and environment metadata
• results.csv: epoch-level training history
• args.yaml: Ultralytics training arguments
• training_metrics.png: metric curves across epochs
• per_class_map50_95.png: per-class mAP50-95
• results.png: Ultralytics training summary
• confusion_matrix.png: validation confusion matrix
• confusion_matrix_normalized.png: normalized confusion matrix
• BoxPR_curve.png: precision-recall curve
• BoxP_curve.png: precision curve
• BoxR_curve.png: recall curve
• BoxF1_curve.png: F1 curve

Baseline Comparison Protocol

All baseline models use:

• the same ACDC train and validation split
• the same image size
• the same batch size
• the same optimizer configuration
• the same augmentation policy
• the same seed
• the same training budget
• the same post-training validation protocol

Hardware

• GPU: Tesla T4
• GPU memory: 14.56 GB
• PyTorch: 2.11.0+cu128
• Ultralytics: 8.4.160