Production-Grade Spatial Understanding with Gemini 2.5 Pro: Bounding Boxes and Segmentation Done Right

Working with Segmentation Masks

Bounding boxes are rectangles, but for many use cases — clean product cutouts, irregular shapes — you want the actual silhouette. Gemini 2.5 Pro can return segmentation masks too, and they're the right tool when a rectangle simply isn't precise enough.

Mask Format and Base64 PNG Decoding

Gemini returns each mask as a small PNG sized to fit the bounding box, embedded as a base64 string. To use it on the original image, you must decode it, resize it to the box's pixel dimensions, and paste it onto a full-size canvas. Skip that resize step and you'll see a tiny ghost mask floating in the upper-left of your image.

import base64
from PIL import Image
from io import BytesIO
 
class Segment(BaseModel):
    label: str
    box_2d: List[int]
    mask: str = Field(description="data:image/png;base64,... mask")
 
class SegmentationResult(BaseModel):
    segments: List[Segment]
 
def segment_objects(image_path: str, target: str) -> SegmentationResult:
    client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])
    image = Image.open(image_path)
 
    prompt = (
        f"For each instance of '{target}' in the image, return its bounding box "
        "and a segmentation mask. Each mask must be a data:image/png;base64,... "
        "PNG sized to fit its box_2d. Return an empty segments array if nothing "
        "matches."
    )
    resp = client.models.generate_content(
        model="gemini-2.5-pro",
        contents=[prompt, image],
        config=types.GenerateContentConfig(
            response_mime_type="application/json",
            response_schema=SegmentationResult,
            temperature=0.1,
        ),
    )
    return SegmentationResult.model_validate_json(resp.text)
 
def decode_mask_to_full_image(seg: Segment, img_w: int, img_h: int) -> Image.Image:
    """Expand the box-fitted base64 mask into a full-image binary mask."""
    b64 = seg.mask.split(",", 1)[1] if "," in seg.mask else seg.mask
    box_png = Image.open(BytesIO(base64.b64decode(b64))).convert("L")
 
    x1, y1, x2, y2 = to_pixel_box(seg.box_2d, img_w, img_h)
    box_w, box_h = max(1, x2 - x1), max(1, y2 - y1)
 
    # The small mask fits the box — resize to the box's pixel size
    box_png = box_png.resize((box_w, box_h), Image.NEAREST)
 
    # Compose onto a black, full-image canvas
    full = Image.new("L", (img_w, img_h), 0)
    full.paste(box_png, (x1, y1))
    return full

Image.NEAREST is intentional. Masks are effectively binary (or grayscale that you'll threshold), and bilinear interpolation introduces gray pixels along edges that wobble after thresholding. If you need anti-aliased composites, do the threshold first and add edge softening as a final step.

In product apps you typically use the mask to crop the object cleanly, then push the cutout into a thumbnail or a search index. The "snap a photo, send a clean cutout to the search backend" loop is exactly the kind of workflow that this format unlocks.

One more practical tip: keep the original image and the mask side by side in storage when you persist a result. A mask without its origin frame is hard to reuse — re-rendering, debugging detection regressions, or training a future custom detector all become much easier when you can recompose the cutout from the source image at any time.

Four Checks That Suppress Hallucinated Coordinates

This is the single most important section for production. Gemini will, sometimes, confidently fabricate coordinates for objects that are not actually in the image. If you ship that to users, an "AI shopping assistant" tells them about products that don't exist on the shelf — credibility is gone in one shot.

The defense I use has four layers.

Check 1: Confidence floor. Ask the model for confidence in the schema and discard anything under 0.5. The numbers aren't calibrated probabilities, but in practice 0.95-class detections behave very differently from 0.55-class ones.

Check 2: Sanity bounds on box size. A box covering 99% of the frame, or 5×5 pixels of nothing, is almost always noise. I drop anything with area_ratio < 0.001 or area_ratio > 0.95.

Check 3: Consistency across runs. Run detection two or three times with slightly different temperatures and only keep boxes that overlap (IoU ≥ 0.5) across runs. Cost goes up, but for accuracy-critical pipelines this works.

Check 4: Re-verification with the model itself. Send the cropped detection back and ask "does this image actually show X — yes or no?" Doubles your bill, but for medical or financial extraction it can be the right trade.

def iou(a: List[int], b: List[int]) -> float:
    """IoU between [ymin, xmin, ymax, xmax] boxes."""
    ay1, ax1, ay2, ax2 = a
    by1, bx1, by2, bx2 = b
    iy1, ix1 = max(ay1, by1), max(ax1, bx1)
    iy2, ix2 = min(ay2, by2), min(ax2, bx2)
    iw, ih = max(0, ix2 - ix1), max(0, iy2 - iy1)
    inter = iw * ih
    if inter == 0:
        return 0.0
    area_a = (ay2 - ay1) * (ax2 - ax1)
    area_b = (by2 - by1) * (bx2 - bx1)
    return inter / (area_a + area_b - inter)
 
def filter_hallucinations(
    detections: List[Detection],
    img_w: int, img_h: int,
    min_conf: float = 0.5,
) -> List[Detection]:
    out = []
    for d in detections:
        if d.confidence < min_conf:
            continue
        x1, y1, x2, y2 = to_pixel_box(d.box_2d, img_w, img_h)
        bw, bh = x2 - x1, y2 - y1
        if bw <= 0 or bh <= 0:
            continue
        ratio = (bw * bh) / float(img_w * img_h)
        if ratio < 0.001 or ratio > 0.95:
            continue
        out.append(d)
    return out
 
def cross_check(detections_run1, detections_run2, iou_threshold: float = 0.5):
    """Keep only detections that match across two runs (same label, IoU >= threshold)."""
    consistent = []
    for d1 in detections_run1:
        for d2 in detections_run2:
            if d1.label == d2.label and iou(d1.box_2d, d2.box_2d) >= iou_threshold:
                consistent.append(d1)
                break
    return consistent

In my own tests, this combined filter dropped "plausible-but-fake" detections from roughly one or two per ten down to nearly zero, at the cost of a few percent recall. That trade-off is almost always worth it: in product UX, "showed a wrong thing" hurts trust far more than "missed something."

Stabilizing Large Images and Many Objects

Real photos often run 4000×3000 pixels or more. Sending that straight to Gemini wastes tokens and seems to nudge spatial precision down (the docs don't quantify this, but my experiments do). Resize the long edge to 1024–1536 first.

def smart_resize(img: Image.Image, max_side: int = 1280) -> Image.Image:
    w, h = img.size
    longest = max(w, h)
    if longest <= max_side:
        return img
    scale = max_side / longest
    return img.resize((round(w * scale), round(h * scale)), Image.LANCZOS)

Because box_2d is normalized 0–1000, you can resize the input freely and still convert back to original pixels with to_pixel_box. That separation between "input size for the model" and "display size for the user" is one of the cleaner aspects of this API.

When the count of target objects gets large (think 100+), Gemini may truncate its output. The honest answer is to tile the image into 2×2 or 3×3 tiles, run detection per tile, and merge with overlap-aware deduplication. In practice, most real apps don't need that extreme scale — instead, ask for "the top N most prominent" and accept the limit.

Another stabilizing trick is to anchor the prompt with the expected image scale. If you tell Gemini the image is "a wide-angle shot of a supermarket aisle" or "a vertical phone photo of a single receipt," the model's spatial reasoning improves measurably. The model has the metadata, but spelling it out in prose helps. I keep a small library of "scene priors" that I prepend per workflow, and that alone tends to lift detection quality more than tweaking model parameters does.

Designing Automatic Fallback to a Dedicated Detector

In production, Gemini will sometimes fail (rate limit, transient 5xx, zero detections). Your app can't stop. The fallback I run on Cloudflare Workers looks like this.

// Cloudflare Workers
import { GoogleGenAI } from "@google/genai";
 
interface DetectInput {
  imageUrl: string;
  target: string;
}
 
interface Detection {
  label: string;
  box_2d: [number, number, number, number];
  confidence: number;
  source: "gemini" | "yolo";
}
 
const SOFT_FAIL_PATTERNS = [
  /quota/i,
  /rate.?limit/i,
  /server.?error/i,
];
 
async function detectWithGemini(input: DetectInput, env: Env): Promise<Detection[]> {
  const ai = new GoogleGenAI({ apiKey: env.GEMINI_API_KEY });
  const imageBytes = await (await fetch(input.imageUrl)).arrayBuffer();
  const result = await ai.models.generateContent({
    model: "gemini-2.5-pro",
    contents: [
      {
        role: "user",
        parts: [
          {
            text: `Detect '${input.target}' in this image. box_2d is [ymin,xmin,ymax,xmax] (0-1000).`,
          },
          {
            inlineData: {
              mimeType: "image/jpeg",
              data: btoa(String.fromCharCode(...new Uint8Array(imageBytes))),
            },
          },
        ],
      },
    ],
    config: {
      responseMimeType: "application/json",
      responseSchema: {
        type: "object",
        properties: {
          detections: {
            type: "array",
            items: {
              type: "object",
              properties: {
                label: { type: "string" },
                box_2d: { type: "array", items: { type: "integer" } },
                confidence: { type: "number" },
              },
              required: ["label", "box_2d", "confidence"],
            },
          },
        },
        required: ["detections"],
      },
      temperature: 0.1,
    },
  });
  const parsed = JSON.parse(result.text ?? "{\"detections\":[]}");
  return parsed.detections.map((d: Detection) => ({ ...d, source: "gemini" as const }));
}
 
async function detectWithYolo(input: DetectInput, env: Env): Promise<Detection[]> {
  // Hosted YOLO endpoint (Replicate, Modal, your own GPU box, etc.)
  const r = await fetch(env.YOLO_ENDPOINT, {
    method: "POST",
    headers: { "Authorization": `Bearer ${env.YOLO_TOKEN}` },
    body: JSON.stringify({ image_url: input.imageUrl, target: input.target }),
  });
  if (!r.ok) return [];
  const json = await r.json<{ detections: Detection[] }>();
  return json.detections.map((d) => ({ ...d, source: "yolo" }));
}
 
export async function detect(input: DetectInput, env: Env): Promise<Detection[]> {
  try {
    const dets = await detectWithGemini(input, env);
    if (dets.length > 0) return dets;
    // Don't trust an empty result blindly — confirm with YOLO
    const fallback = await detectWithYolo(input, env);
    return fallback;
  } catch (e: any) {
    const msg = String(e?.message ?? e);
    const recoverable = SOFT_FAIL_PATTERNS.some((p) => p.test(msg));
    if (recoverable) {
      console.warn("Gemini soft-fail, falling back to YOLO:", msg);
      return detectWithYolo(input, env);
    }
    throw e;
  }
}

The non-obvious bit is "if Gemini returns zero, ask YOLO anyway." Gemini's calibration leans toward false negatives when the prompt requests caution, so empty results are not as confidence-inducing as they look. YOLO covering for that lifts recall back up. If Gemini says yes and YOLO says no, you have a soft signal to lower the surfaced confidence or hold the result until further verification.

A Testing Strategy That Catches Detection Regressions

A subtle problem with shipping Gemini-based detection is that the model itself can shift behavior under your feet — a new model version, an updated prompt, or even seasonal load changes can move detection quality. You need a regression suite, but a traditional unit test framework feels awkward here.

What works for me is a fixture-driven snapshot test: keep a folder of representative images, each with an expected set of detections (label + approximate box), and a test that runs detection and compares to the expected set with IoU as the matching criterion.

import json
from pathlib import Path
 
def evaluate(predicted, expected, iou_thr=0.4):
    # Returns (precision, recall) treating IoU >= iou_thr + same label as a match.
    matched_pred = set()
    matched_exp = set()
    for i, e in enumerate(expected):
        for j, p in enumerate(predicted):
            if j in matched_pred:
                continue
            if e["label"] != p["label"]:
                continue
            if iou(e["box_2d"], p["box_2d"]) >= iou_thr:
                matched_pred.add(j)
                matched_exp.add(i)
                break
    precision = len(matched_pred) / max(1, len(predicted))
    recall = len(matched_exp) / max(1, len(expected))
    return precision, recall
 
def run_regression_suite(fixtures_dir: str, threshold_p: float = 0.7, threshold_r: float = 0.7):
    failures = []
    for case in Path(fixtures_dir).glob("*.json"):
        spec = json.loads(case.read_text())
        result = detect_objects(spec["image_path"], spec["target"])
        precision, recall = evaluate(
            [d.model_dump() for d in result.detections],
            spec["expected"],
        )
        if precision < threshold_p or recall < threshold_r:
            failures.append({
                "case": case.name,
                "precision": precision,
                "recall": recall,
            })
    return failures

Run this in CI on a small fixed image set. When you bump models or prompts, the suite tells you whether quality moved in the wrong direction before users find out. I would rather pay for a few hundred extra detection calls per release than be debugging a "the AI suddenly stopped finding products" Slack thread on a Friday night.

Cost and Latency Notes from the Field

Concrete numbers shift faster than any article can keep current, so treat these as illustrative rather than authoritative. With Gemini 2.5 Pro on a 1280-pixel image, single-target detection that returns five to ten boxes lands somewhere in the few-cents range and completes in 1–2 seconds end to end. Adding segmentation masks roughly doubles the latency and bumps cost noticeably because the model emits significantly more output tokens.

What that means for product decisions: detection-only is cheap enough to put on every uploaded image, but masks should be opt-in — a "show contour" toggle in the UI, or only requested when the next step actually needs the mask. The latency mix also matters; if your UX shows results inline, prefer detection-only, and if you can hide the wait inside an upload progress sheet, masks become usable.

For very high-volume pipelines, I shift to Gemini 2.5 Flash for first-pass detection and only escalate to 2.5 Pro for difficult cases (low confidence or zero results). The Flash-to-Pro escalator pattern, combined with the YOLO fallback, gives you a three-layer pipeline that balances cost, latency, and accuracy in a way no single model can.

Practical Scenarios for Mobile Camera Apps

Here are a few places this combo earns its keep, drawn from years of running indie mobile apps.

For e-commerce "find this product" features, you can pass the user's shelf photo to Gemini with a natural-language target like "snack bars" or "cleaning sprays," crop the detected region, and feed those crops into your product search index. Gemini's zero-shot semantic targeting is the right tool for arbitrary user queries that you can't preregister with a YOLO class list.

For receipt and invoice OCR, asking Gemini to box specifically the "total amount" region and then OCR-ing only that high-resolution crop gave me materially better accuracy than running OCR over the full image and post-processing. The two-stage pipeline (box → OCR) is worth the small extra latency.

For wallpaper, mood, and lifestyle apps — where "the impressive subject in this photo" is the kind of vague concept users actually express — Gemini handles requests no purpose-built detector can. Crop, frame, and effect pipelines benefit massively here.

Run the call during a transition moment — the post-shutter loading screen, the upload progress sheet — so 1–2 second latency feels justified rather than wasted. That's the UX seam where this API fits cleanly.

If you're building accessibility features — describing scenes, narrating what's in front of a camera for low-vision users — Gemini's spatial outputs let you say "there is a cup on the left edge, a phone in the center, and a notebook to the right" rather than the generic "this image contains objects." That spatial-aware narration is one of the more meaningful applications I've seen for this API, and it's the kind of feature that's nearly impossible to build with a fixed-class detector.

Common Pitfalls

These are the traps that have cost me real time.

Pitfall 1: Treating coordinates as (x, y). Gemini returns [ymin, xmin, ymax, xmax]. If you read box_2d[0] as X, your boxes look rotated or escape the image. Centralize the conversion in a helper like to_pixel_box so the mistake can only happen once.

Pitfall 2: Pasting masks at the wrong scale. The base64 PNG is sized to the box, not the full image. Without decode_mask_to_full_image, you'll see a small ghost in the corner. Resize-to-box-then-composite is mandatory.

Pitfall 3: Setting temperature to zero. Tempting in the name of "deterministic," but it makes detection so conservative that users complain it's blind. 0.1–0.2 is the sweet spot.

Pitfall 4: Trusting the confidence as a probability. It's an ordering signal, not a calibrated probability. Use it to threshold, never to make statistical claims to users.

Pitfall 5: Sending huge images raw. Past 4000-ish pixels, you waste tokens and lose a touch of precision. Apply smart_resize to 1024–1536 by default.

Pitfall 6: Ignoring "the object isn't there" cases. Gemini sometimes fabricates rather than admit nothing matches. Combine an explicit "return empty if absent" instruction with the filter_hallucinations pass downstream.

Pitfall 7: Hammering the API with concurrent requests. Naive UI wiring ("detect on every selection") triggers rate limits when users tap fast. Debounce and use AbortController to cancel obsolete in-flight calls. Both UX and cost benefit.

Pitfall 8: Forgetting to log the raw model output. Once you start filtering and post-processing, you lose the ability to diagnose detection regressions later. Persist the raw JSON response in object storage with a hash of the input image as the key. When users report "the AI marked the wrong thing," you can replay exactly what the model returned without paying for a re-detection.

For the surrounding production concerns, the Gemini API rate-limiting and quota-management guide and the streaming-cancel + AbortController guide round out the playbook nicely.

Operational Concerns You'll Hit on Day Two

After the first release, several smaller issues tend to surface. Documenting them up front saves the inevitable second sprint of "we didn't think about that" patches.

Image privacy and retention. Gemini's terms allow image data to be processed for the request, but you should still treat user images as sensitive. Keep them in your own storage with explicit retention windows, never hand-roll API requests that log raw image data, and strip EXIF metadata before sending. I scrub GPS coordinates and device serial numbers from EXIF specifically, because nothing surfaces a privacy issue faster than an indie app's image pipeline accidentally surfacing a user's home address.

Multi-language UI implications. If your app serves Japanese, English, and Spanish users, each will phrase the same target differently — "PET bottles," "ペットボトル," "botellas de plástico." You can either translate the user's input to a canonical language before prompting Gemini, or feed the original phrase straight to the model. In my experience, English target labels behave the most consistently across image content, but native phrases work fine for clearly visible objects. The painful cases are loanwords and brand names — those benefit from a small translation step.

Model version pinning. When Gemini publishes a new model alias, your detections may change subtly even if your code did not. I pin a specific model version in my production config and only roll forward after running the regression suite from the previous section. The alternative — silently inheriting the latest model — has bitten me twice with quality regressions that took days to attribute.

On-device caching. If the same image is uploaded multiple times (think app users retrying after a network failure), you don't need to re-detect. A simple hash of the image bytes plus the target string makes a clean cache key. I keep this in Cloudflare KV with a 24-hour TTL, which removes a meaningful chunk of cost in apps where users retry frequently.

Observability. Beyond per-call cost and latency, I emit a structured event for every detection: target string, image hash, model used, raw box count, filtered box count, confidence histogram, and which fallback path fired. That log lets me answer questions like "is detection getting worse over time?" or "how often does the YOLO fallback save us?" without a one-off investigation. The cost of logging this is negligible; the cost of not having it is that you debug detection regressions blind.

Wrapping Up — Your Next Step

Thanks for reading this far. If I can ask one specific action: take an image you actually have lying around and run detect_objects on it. The 0–1000, Y-first coordinate convention is the kind of thing that only sticks once you've handled it once with your own eyes.

Then, before you ship, layer in filter_hallucinations, cross_check, and the YOLO fallback. Gemini's spatial understanding isn't a replacement for dedicated detectors — it's a complement that shines on conceptual queries. Wiring both into one pipeline is what makes this practical in real apps.