75 lines
3.6 KiB
Python
75 lines
3.6 KiB
Python
#!/usr/bin/env python3
|
||
"""开放词表分割(texts → mask + box)—— 走 prod_server HTTP 服务(/api/v1/segment,texts 分支)。
|
||
|
||
结构镜像 demos/00_text_grounding.py:给一段/多段开放词表文本,检测并分割图中**所有匹配实例**,
|
||
输出每个实例的 mask、包围框、置信度分数。唯一差别——推理不调本地 `proc.set_text_prompt`,
|
||
而是 POST 给服务;服务按短语循环 grounding,返回跨短语实例(每条带 `text/box/score`)。
|
||
|
||
渲染/IO/路径全部复用 demos/_common(与 demos 同源,mask+框+`#编号`+分数同款);HTTP 走 _client_common。
|
||
`texts` 是与 `points`/`bboxes` 平级的第三个 prompt 字段(list)。在 sam3 conda env 下运行(同 demos)。
|
||
|
||
python3 prod_server/client_segment_text.py
|
||
python3 prod_server/client_segment_text.py --input assets/inputs/groceries.jpg --texts "a plastic bag" "a bottle" --threshold 0.5
|
||
"""
|
||
import argparse
|
||
import os
|
||
import sys
|
||
|
||
import numpy as np
|
||
|
||
HERE = os.path.dirname(os.path.abspath(__file__))
|
||
sys.path.insert(0, os.path.join(HERE, "..", "demos")) # demos/_common
|
||
sys.path.insert(0, HERE) # _client_common
|
||
from _common import load_image, overlay_masks, save_image, in_path, out_path
|
||
from _client_common import file_to_data_uri, post, polygons_to_mask
|
||
|
||
|
||
def run(input_path, texts, output_path, threshold=0.5):
|
||
img = load_image(input_path)
|
||
|
||
# 1) 推理走服务:POST /api/v1/segment(texts 分支,吃开放词表短语列表)
|
||
payload = {
|
||
"image": file_to_data_uri(input_path),
|
||
"model": "SAM-3",
|
||
"texts": list(texts),
|
||
"threshold": threshold,
|
||
}
|
||
resp = post("/api/v1/segment", payload)
|
||
print(f"[texts] status={resp['status']} message={resp['message']!r} texts={texts} threshold={threshold}")
|
||
|
||
data = resp.get("data") or []
|
||
if not data:
|
||
print("[texts] ⚠ 服务未返回实例(短语可能未命中 / threshold 过高)")
|
||
save_image(img, output_path)
|
||
return
|
||
|
||
# 2) 还原掩码并叠加框+分数(demos 同款 overlay_masks;boxes 用 xyxy)
|
||
w, h = img.size
|
||
masks, boxes, scores = [], [], []
|
||
per_phrase = {}
|
||
for item in data:
|
||
polys = item.get("polygons") or []
|
||
x, y, bw, bh = item["box"] # xywh
|
||
masks.append(polygons_to_mask(polys, w, h))
|
||
boxes.append([x, y, x + bw, y + bh]) # xywh -> xyxy
|
||
scores.append(float(item["score"]))
|
||
per_phrase[item["text"]] = per_phrase.get(item["text"], 0) + 1
|
||
masks = np.stack(masks) # (N,H,W)
|
||
boxes = np.array(boxes, dtype=float) # (N,4) xyxy
|
||
scores = np.array(scores, dtype=float) # (N,)
|
||
|
||
print(f"[texts] 共 {len(data)} 个实例:" + ", ".join(f"{k!r}->{v}" for k, v in per_phrase.items()))
|
||
save_image(overlay_masks(img, masks, boxes=boxes, scores=scores), output_path)
|
||
print(f"[texts] 结果图: {output_path}")
|
||
|
||
|
||
if __name__ == "__main__":
|
||
ap = argparse.ArgumentParser(description="开放词表分割(走 prod_server 服务;镜像 demos 00)")
|
||
ap.add_argument("--input", default=in_path("truck.jpg"), help="输入图(默认 demos 同图)")
|
||
ap.add_argument("--texts", nargs="+", default=["a truck"], metavar="PHRASE",
|
||
help="开放词表短语列表(与 points/bboxes 平级;默认 ['a truck'])")
|
||
ap.add_argument("--threshold", type=float, default=0.5, help="置信度阈值(默认 0.5;越低实例越多)")
|
||
ap.add_argument("--output", default=out_path("server_00_truck.jpg"), help="结果图输出")
|
||
args = ap.parse_args()
|
||
run(args.input, args.texts, args.output, args.threshold)
|