TBX11K SSD300 — Tuberculosis Object Detection

Per-class Performance (mAP@0.5 Threshold)

Global Summary Scores (900 Isolated Test Samples)

• Global Test mAP@0.5 : 0.3892 (38.9%)
• Global Test COCO mAP@0.5:0.95 : 0.2669 (26.7%)
• Global Test mAP@0.75 : 0.1446 (14.5%)

Classes & Target Deployment Indices

• 0: __background__ (Strict PyTorch Detection Placeholder)
• 1: healthy (Includes clear lungs and latent tuberculosis signals)
• 2: sick_non_tb (Other pulmonary conditions / non-TB sickness)
• 3: active_tb (Active Tuberculosis visual lesions)

Model Configuration

• Input Resolution: 300x300 pixels
• Optimization Engine: AdamW + 5-Epoch Linear Warmup & Cosine Annealing
• Oversampling Strategy: Square-Root Weighted Random Sampler targeting Active TB regions