File size: 9,316 Bytes
c10636a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 | """Count visible objects with schema-constrained vision output."""
from __future__ import annotations
import math
import os
import tempfile
from dataclasses import dataclass, field
from typing import Any, Dict, Optional
from eviscene.integrations.gemini import RUNTIME_METRICS_KEY, exception_runtime_metrics
from eviscene.tools.base import BaseTool
COUNT_PROMPT_TEMPLATE = """\
You are the object-counting module of an image verification system.
Count only the visible instances of the requested target.
Target: {target_object}
Rules:
1. Inspect the full supplied image or crop systematically.
2. Count a partially occluded instance only when its presence is visually supported.
3. Return zero when no target instance is visible.
4. Do not infer instances outside the frame or hidden behind objects.
5. Keep the explanation literal and concise.
6. Return exactly one JSON object matching the response schema.
"""
COUNT_RESPONSE_SCHEMA: Dict[str, Any] = {
"type": "object",
"properties": {
"target": {"type": "string", "maxLength": 200},
"count": {"type": "integer", "minimum": 0},
"confidence": {"type": "number", "minimum": 0.0, "maximum": 1.0},
"details": {"type": "string", "maxLength": 800},
"locations": {
"type": "array",
"items": {"type": "string", "maxLength": 240},
"maxItems": 50,
},
},
}
@dataclass
class CountObjectsTool(BaseTool):
"""Count a requested visible object in the image or a normalized crop."""
name: str = "count_objects"
description: str = (
"Count a specific visible object in the image or in a normalized crop. "
"Provide a precise target such as 'fingers on the left hand' or "
"'people in the background'. Returns a count, confidence, and locations."
)
parameters: dict = field(
default_factory=lambda: {
"type": "object",
"properties": {
"image_input": {
"type": "string",
"description": "Local path to the image file.",
},
"target_object": {
"type": "string",
"description": "A precise description of what to count.",
},
"bbox": {
"type": "array",
"items": {"type": "number"},
"minItems": 4,
"maxItems": 4,
"description": (
"Optional normalized crop [x1, y1, x2, y2], with every "
"coordinate in [0, 1]."
),
},
"visual_question_id": {"type": "string"},
"source_evidence_id": {"type": "string"},
"source_discovery_id": {"type": "string"},
"expected_property": {"type": "string"},
},
"required": ["image_input", "target_object"],
}
)
client: Optional[Any] = field(default=None, repr=False)
provider: str = "gemini"
model_name: str = "gemini-3.7-flash"
def _get_client(self) -> Any:
if self.client is None:
from eviscene.integrations.vlm.factory import build_vlm_client
self.client = build_vlm_client(
provider=self.provider,
model_name=self.model_name,
)
return self.client
def call(self, params: Dict[str, Any]) -> Dict[str, Any]:
image_input = str(params.get("image_input", "")).strip()
target_object = str(params.get("target_object", "")).strip()
if not image_input:
return {"status": "error", "error": "image_input is required."}
if not target_object:
return {"status": "error", "error": "target_object is required."}
actual_image = image_input
temporary_path: Optional[str] = None
bbox = params.get("bbox")
if bbox is not None:
try:
coordinates = self._validate_bbox(bbox)
actual_image, temporary_path = self._crop(image_input, coordinates)
except Exception as exc:
return {
"status": "error",
"error": f"Requested count region could not be cropped: {exc}",
}
try:
parsed = self._get_client().create_image_json(
system_prompt=COUNT_PROMPT_TEMPLATE.format(
target_object=target_object,
),
user_text=f"Count the visible instances of: {target_object}",
image_input=actual_image,
max_tokens=1000,
model_name=self.model_name,
response_schema=COUNT_RESPONSE_SCHEMA,
)
except Exception as exc:
error = {
"status": "error",
"error": f"Counting failed: {type(exc).__name__}: {exc or '<no message>'}",
}
metrics = exception_runtime_metrics(exc)
if metrics:
error[RUNTIME_METRICS_KEY] = metrics
return error
finally:
if temporary_path:
try:
os.remove(temporary_path)
except OSError:
pass
result = self._validate_response(parsed, target_object)
runtime_metrics = parsed.get(RUNTIME_METRICS_KEY)
if runtime_metrics:
result[RUNTIME_METRICS_KEY] = runtime_metrics
return result
@staticmethod
def _validate_bbox(value: Any) -> tuple[float, float, float, float]:
if not isinstance(value, (list, tuple)) or len(value) != 4:
raise ValueError("bbox must contain exactly four normalized coordinates")
if any(isinstance(item, bool) or not isinstance(item, (int, float)) for item in value):
raise ValueError("bbox coordinates must be numeric")
coordinates = tuple(float(item) for item in value)
if any(not math.isfinite(item) or item < 0.0 or item > 1.0 for item in coordinates):
raise ValueError("bbox coordinates must be finite values in [0, 1]")
x1, y1, x2, y2 = coordinates
if x1 >= x2 or y1 >= y2:
raise ValueError("bbox must satisfy x1 < x2 and y1 < y2")
return x1, y1, x2, y2
@staticmethod
def _crop(
image_input: str,
bbox: tuple[float, float, float, float],
) -> tuple[str, str]:
from PIL import Image
with Image.open(image_input) as image:
width, height = image.size
x1, y1, x2, y2 = bbox
pixel_box = (
max(0, min(int(x1 * width), width - 1)),
max(0, min(int(y1 * height), height - 1)),
max(1, min(int(math.ceil(x2 * width)), width)),
max(1, min(int(math.ceil(y2 * height)), height)),
)
if pixel_box[0] >= pixel_box[2] or pixel_box[1] >= pixel_box[3]:
raise ValueError("bbox resolves to an empty crop")
cropped = image.crop(pixel_box)
handle, temporary_path = tempfile.mkstemp(prefix="count_objects_", suffix=".png")
os.close(handle)
try:
cropped.save(temporary_path)
except Exception:
try:
os.remove(temporary_path)
except OSError:
pass
raise
return temporary_path, temporary_path
@staticmethod
def _validate_response(parsed: Any, target_object: str) -> Dict[str, Any]:
if not isinstance(parsed, dict):
return {"status": "error", "error": "Counting response must be a JSON object."}
count = parsed.get("count")
if isinstance(count, bool) or not isinstance(count, int) or count < 0:
return {
"status": "error",
"error": "Counting response must contain a non-negative integer 'count'.",
}
confidence = parsed.get("confidence")
if isinstance(confidence, bool) or not isinstance(confidence, (int, float)):
return {
"status": "error",
"error": "Counting response must contain numeric 'confidence'.",
}
confidence = float(confidence)
if not math.isfinite(confidence) or not 0.0 <= confidence <= 1.0:
return {
"status": "error",
"error": "Counting confidence must be a finite value in [0, 1].",
}
details = parsed.get("details")
locations = parsed.get("locations")
if not isinstance(details, str):
return {"status": "error", "error": "Counting details must be a string."}
if not isinstance(locations, list) or any(
not isinstance(location, str) for location in locations
):
return {"status": "error", "error": "Counting locations must be an array of strings."}
return {
"status": "success",
"target": target_object,
"count": count,
"confidence": confidence,
"details": details.strip(),
"locations": [location.strip() for location in locations],
}
|