arc-vision

Vision models for arc-vision — the autonomous drone vision pipeline.

Files

File	Size	Description
yolov11s.hef	16 MB	YOLO11s object detection (640×640, 80 COCO classes, Hailo-10H)
clip_vit_b_16_text_encoder.safetensors	254 MB	CLIP ViT-B/16 text encoder weights for target designation
clip_tokenizer.json	3.6 MB	CLIP tokenizer
clip_vit_b_16_image_encoder.hef	76 MB	CLIP ViT-B/16 image encoder compiled for Hailo-10H NPU

Architecture

• Detection: YOLO11s on Hailo-10H NPU (~46 FPS)
• Target designation: CLIP text encoder (CPU, candle) + image encoder (NPU, ~53 FPS) for zero-shot matching
• Re-ID: CLIP image embeddings for cross-frame track association

Describe a target in natural language ("silver car", "person in red jacket") and the pipeline matches against live camera frames.

Sources

• YOLO11s + CLIP ViT-B/16 image encoder: compiled from Hailo Model Zoo v5.2.0
• CLIP text encoder + tokenizer: exported from openai/clip-vit-base-patch16