🧬 三技术融合 · AI城能力注入

from moondream.onnx_vl import OnnxVL model = OnnxVL.from_path("moondream-0.5b.moon") # caption → "a red rectangle and green rectangle" # query → "The rectangle on the left is red" # detect → 2 rectangles with bbox coords

模型下载命令（待代理）: $ export HTTPS_PROXY=socks5://127.0.0.1:10808 $ python3 -c "from huggingface_hub import hf_hub_download hf_hub_download('whistleroosh/moondream-0.5B','moondream-0.5b.moon', local_dir='~/.hermes/models')"

class GazeSenseChannel: """域灵感知第五通道 — 基于Moondream的视线感知""" def __init__(self, model_path): from moondream.onnx_vl import OnnxVL self.model = OnnxVL.from_path(model_path) def detect_faces(self, image) -> list[Region]: """检测画面中所有人脸""" return self.model.detect(image, "face")["objects"] def compute_gaze(self, image, faces) -> list[GazeVector]: """计算每个人的视线落点""" results = [] for face in faces: # query模型: "where is this person looking?" ans = self.model.query(crop(image, face), "Is this person looking at the center, left, right, up or down?") results.append(GazeVector(face, ans["answer"])) return results def shared_attention(self, gazes) -> list[AttentionCluster]: """共享注意力聚类 — 多人视线交汇点""" return cluster_gaze_intersections(gazes) def attention_heatmap(self, image, gazes) -> Image: """生成注意力热图""" ...