# Core
numpy>=1.24
# Vision-language backbone (CLIP ViT-B/16, OpenAI weights)
torch>=2.0
torchvision>=0.15
open-clip-torch>=2.20
pillow>=9.5
# Fully-predicted condition: open-vocabulary detector (default backend)
transformers>=4.40          # Grounding DINO (IDEA-Research/grounding-dino-tiny)
# Optional alternative detector backend
# ultralytics>=8.1          # YOLO-World (uncomment if using --backend yoloworld)
# Plotting
matplotlib>=3.7
