Object Detector

The MediaPipe Object Detector task lets you detect the presence and location of multiple classes of objects within images or videos. For example, an object detector can locate dogs in an image. This task operates on image data with a machine learning (ML) model, accepting static data or a continuous video stream as input and outputting a list of detection results. Each detection result represents an object that appears within the image or video.

Open in Web DemosEdit in StackBlitz

Quickstart

Install the SDK for your platform, prepare input data, choose a running mode, and run inference.

Web

npmnpm install @mediapipe/tasks-visionWeb demo·Edit in StackBlitz·Setup guide

Image

IMAGEStateless · single frameSynchronous call on independent still images. Resets temporal tracking state on every call.
import { FilesetResolver, ObjectDetector } from "@mediapipe/tasks-vision";

const vision = await FilesetResolver.forVisionTasks(
  "https://cdn.jsdelivr.net/npm/@mediapipe/tasks-vision@latest/wasm"
);
const objectDetector = await ObjectDetector.createFromOptions(vision, {
  baseOptions: { modelAssetPath: "efficientdet_lite0.tflite", delegate: "GPU" },
  runningMode: "IMAGE",
  scoreThreshold: 0.5,
  maxResults: 5,
});

// Synchronous single-frame call (stateless across images)
const image = document.getElementById("image") as HTMLImageElement;
const result = objectDetector.detect(image);

Video

VIDEOEvery frame · no dropsSynchronous tracking across decoded video frames with monotonic timestamps. Processes every frame in order and never drops frames.
import { FilesetResolver, ObjectDetector } from "@mediapipe/tasks-vision";

const vision = await FilesetResolver.forVisionTasks(
  "https://cdn.jsdelivr.net/npm/@mediapipe/tasks-vision@latest/wasm"
);
const objectDetector = await ObjectDetector.createFromOptions(vision, {
  baseOptions: { modelAssetPath: "efficientdet_lite0.tflite", delegate: "GPU" },
  runningMode: "VIDEO",
  scoreThreshold: 0.5,
  maxResults: 5,
});

// Synchronous across decoded frames with monotonic timestamps (never drops frames)
for (const { frame, timestampMs } of decodedVideoFrames) {
  const result = objectDetector.detectForVideo(frame, timestampMs);
}

Livestream

VIDEO + rAFLive camera · latest frameWeb tasks run synchronously—pair runningMode: "VIDEO" with requestAnimationFrame so each paint inspects only the latest camera frame (skipping intermediate frames when busy).
import { FilesetResolver, ObjectDetector } from "@mediapipe/tasks-vision";

const vision = await FilesetResolver.forVisionTasks(
  "https://cdn.jsdelivr.net/npm/@mediapipe/tasks-vision@latest/wasm"
);
// On Web, live camera loops pair runningMode: "VIDEO" with requestAnimationFrame
const objectDetector = await ObjectDetector.createFromOptions(vision, {
  baseOptions: { modelAssetPath: "efficientdet_lite0.tflite", delegate: "GPU" },
  runningMode: "VIDEO",
  scoreThreshold: 0.5,
  maxResults: 5,
});

const video = document.getElementById("webcam") as HTMLVideoElement;
video.srcObject = await navigator.mediaDevices.getUserMedia({ video: true });
await video.play();
let lastVideoTime = -1;

function renderLoop() {
  if (video.readyState >= 2 && video.currentTime !== lastVideoTime) {
    lastVideoTime = video.currentTime;
    const result = objectDetector.detectForVideo(video, performance.now());
    console.log(result.detections);
  }
  requestAnimationFrame(renderLoop);
}
renderLoop();

Android (Kotlin)

Gradleimplementation("com.google.mediapipe:tasks-vision:latest.release")Sample code·Setup guide

Image

IMAGEStateless · single frameSynchronous call on independent still images. Resets temporal tracking state on every call.
import com.google.mediapipe.framework.image.BitmapImageBuilder
import com.google.mediapipe.framework.image.MPImage
import com.google.mediapipe.tasks.core.BaseOptions
import com.google.mediapipe.tasks.vision.core.RunningMode
import com.google.mediapipe.tasks.vision.objectdetector.ObjectDetector

val baseOptions = BaseOptions.builder()
    .setModelAssetPath("efficientdet_lite0.tflite")
    .build()
val options = ObjectDetector.ObjectDetectorOptions.builder()
    .setBaseOptions(baseOptions)
    .setRunningMode(RunningMode.IMAGE)
    .setScoreThreshold(0.5f)
    .setMaxResults(5)
    .build()

val objectDetector = ObjectDetector.createFromOptions(context, options)

// Synchronous single-image call (no temporal tracking across calls)
val mpImage = BitmapImageBuilder(bitmap).build()
val result = objectDetector.detect(mpImage)

Video

VIDEOEvery frame · no dropsSynchronous tracking across decoded video frames with monotonic timestamps. Processes every frame in order and never drops frames.
import com.google.mediapipe.framework.image.BitmapImageBuilder
import com.google.mediapipe.framework.image.MPImage
import com.google.mediapipe.tasks.core.BaseOptions
import com.google.mediapipe.tasks.vision.core.RunningMode
import com.google.mediapipe.tasks.vision.objectdetector.ObjectDetector

val baseOptions = BaseOptions.builder()
    .setModelAssetPath("efficientdet_lite0.tflite")
    .build()
val options = ObjectDetector.ObjectDetectorOptions.builder()
    .setBaseOptions(baseOptions)
    .setRunningMode(RunningMode.VIDEO)
    .setScoreThreshold(0.5f)
    .setMaxResults(5)
    .build()

val objectDetector = ObjectDetector.createFromOptions(context, options)

// Synchronous: processes every frame in order with monotonic timestamps (no frame drops)
val mpImage = BitmapImageBuilder(frameBitmap).build()
val result = objectDetector.detectForVideo(mpImage, frameTimestampMs)

Livestream

LIVE_STREAMAsync · drops frames when busyAsynchronous camera pipeline with a result callback. Automatically drops incoming frames when busy so results stay current with low latency.
import android.os.SystemClock
import android.util.Log
import com.google.mediapipe.framework.image.BitmapImageBuilder
import com.google.mediapipe.framework.image.MPImage
import com.google.mediapipe.tasks.core.BaseOptions
import com.google.mediapipe.tasks.vision.core.RunningMode
import com.google.mediapipe.tasks.vision.objectdetector.ObjectDetector

val baseOptions = BaseOptions.builder()
    .setModelAssetPath("efficientdet_lite0.tflite")
    .build()
val options = ObjectDetector.ObjectDetectorOptions.builder()
    .setBaseOptions(baseOptions)
    .setRunningMode(RunningMode.LIVE_STREAM)
    .setScoreThreshold(0.5f)
    .setMaxResults(5)
    .setResultListener { result, _ -> val detections = result.detections() }
    .setErrorListener { error -> Log.e("MediaPipe", error.message.orEmpty()) }
    .build()

val objectDetector = ObjectDetector.createFromOptions(context, options)

// Async: automatically drops incoming camera frames when busy so results stay current
objectDetector.detectAsync(mpImage, SystemClock.uptimeMillis())

Android (Java)

Gradleimplementation 'com.google.mediapipe:tasks-vision:latest.release'Sample code·Setup guide

Image

IMAGEStateless · single frameSynchronous call on independent still images. Resets temporal tracking state on every call.
import com.google.mediapipe.framework.image.BitmapImageBuilder;
import com.google.mediapipe.framework.image.MPImage;
import com.google.mediapipe.tasks.core.BaseOptions;
import com.google.mediapipe.tasks.vision.core.RunningMode;
import com.google.mediapipe.tasks.vision.objectdetector.ObjectDetector;
import com.google.mediapipe.tasks.vision.objectdetector.ObjectDetectorResult;

BaseOptions baseOptions = BaseOptions.builder()
    .setModelAssetPath("efficientdet_lite0.tflite")
    .build();
ObjectDetector.ObjectDetectorOptions options = ObjectDetector.ObjectDetectorOptions.builder()
    .setBaseOptions(baseOptions)
    .setRunningMode(RunningMode.IMAGE)
    .setScoreThreshold(0.5f)
    .setMaxResults(5)
    .build();

ObjectDetector objectDetector = ObjectDetector.createFromOptions(context, options);

// Synchronous single-image call (no temporal tracking across calls)
MPImage mpImage = new BitmapImageBuilder(bitmap).build();
ObjectDetectorResult result = objectDetector.detect(mpImage);

Video

VIDEOEvery frame · no dropsSynchronous tracking across decoded video frames with monotonic timestamps. Processes every frame in order and never drops frames.
import com.google.mediapipe.framework.image.BitmapImageBuilder;
import com.google.mediapipe.framework.image.MPImage;
import com.google.mediapipe.tasks.core.BaseOptions;
import com.google.mediapipe.tasks.vision.core.RunningMode;
import com.google.mediapipe.tasks.vision.objectdetector.ObjectDetector;
import com.google.mediapipe.tasks.vision.objectdetector.ObjectDetectorResult;

BaseOptions baseOptions = BaseOptions.builder()
    .setModelAssetPath("efficientdet_lite0.tflite")
    .build();
ObjectDetector.ObjectDetectorOptions options = ObjectDetector.ObjectDetectorOptions.builder()
    .setBaseOptions(baseOptions)
    .setRunningMode(RunningMode.VIDEO)
    .setScoreThreshold(0.5f)
    .setMaxResults(5)
    .build();

ObjectDetector objectDetector = ObjectDetector.createFromOptions(context, options);

// Synchronous: processes every frame in order with monotonic timestamps (no frame drops)
MPImage mpImage = new BitmapImageBuilder(frameBitmap).build();
ObjectDetectorResult result = objectDetector.detectForVideo(mpImage, frameTimestampMs);

Livestream

LIVE_STREAMAsync · drops frames when busyAsynchronous camera pipeline with a result callback. Automatically drops incoming frames when busy so results stay current with low latency.
import android.os.SystemClock;
import android.util.Log;
import com.google.mediapipe.framework.image.BitmapImageBuilder;
import com.google.mediapipe.framework.image.MPImage;
import com.google.mediapipe.tasks.core.BaseOptions;
import com.google.mediapipe.tasks.vision.core.RunningMode;
import com.google.mediapipe.tasks.vision.objectdetector.ObjectDetector;

BaseOptions baseOptions = BaseOptions.builder()
    .setModelAssetPath("efficientdet_lite0.tflite")
    .build();
ObjectDetector.ObjectDetectorOptions options = ObjectDetector.ObjectDetectorOptions.builder()
    .setBaseOptions(baseOptions)
    .setRunningMode(RunningMode.LIVE_STREAM)
    .setScoreThreshold(0.5f)
    .setMaxResults(5)
    .setResultListener((result, inputImage) -> { var detections = result.detections(); })
    .setErrorListener(error -> Log.e("MediaPipe", String.valueOf(error.getMessage())))
    .build();

ObjectDetector objectDetector = ObjectDetector.createFromOptions(context, options);

// Async: automatically drops incoming camera frames when busy so results stay current
objectDetector.detectAsync(mpImage, SystemClock.uptimeMillis());

iOS

SwiftPMhttps://github.com/google-ai-edge/mediapipe → MediaPipeTasksVisionPackage.swift·Sample code·Setup guide

Image

IMAGEStateless · single frameSynchronous call on independent still images. Resets temporal tracking state on every call.
import Foundation
import MediaPipeTasksVision

guard let modelPath = Bundle.main.path(forResource: "efficientdet_lite0", ofType: "tflite") else { return }

let options = ObjectDetectorOptions()
options.baseOptions.modelAssetPath = modelPath
options.runningMode = .image
options.scoreThreshold = 0.5
options.maxResults = 5

let objectDetector = try ObjectDetector(options: options)

// Synchronous single-image call (stateless across images)
let mpImage = try MPImage(uiImage: image)
let result = try objectDetector.detect(image: mpImage)

Video

VIDEOEvery frame · no dropsSynchronous tracking across decoded video frames with monotonic timestamps. Processes every frame in order and never drops frames.
import Foundation
import MediaPipeTasksVision

guard let modelPath = Bundle.main.path(forResource: "efficientdet_lite0", ofType: "tflite") else { return }

let options = ObjectDetectorOptions()
options.baseOptions.modelAssetPath = modelPath
options.runningMode = .video
options.scoreThreshold = 0.5
options.maxResults = 5

let objectDetector = try ObjectDetector(options: options)

// Synchronous: processes every frame in order with monotonic timestamps (no frame drops)
let mpImage = try MPImage(pixelBuffer: pixelBuffer)
let result = try objectDetector.detect(videoFrame: mpImage, timestampInMilliseconds: timestampMs)

Livestream

LIVE_STREAMAsync · drops frames when busyAsynchronous camera pipeline with a result callback. Automatically drops incoming frames when busy so results stay current with low latency.
import Foundation
import MediaPipeTasksVision

final class ObjectDetectorStreamService: NSObject {
    private var objectDetector: ObjectDetector?

    func setUp() throws {
        guard let modelPath = Bundle.main.path(
            forResource: "efficientdet_lite0",
            ofType: "tflite"
        ) else { return }

        let options = ObjectDetectorOptions()
        options.baseOptions.modelAssetPath = modelPath
        options.runningMode = .liveStream
        options.scoreThreshold = 0.5
        options.maxResults = 5
        options.objectDetectorLiveStreamDelegate = self

        objectDetector = try ObjectDetector(options: options)
    }

    func detectAsync(mpImage: MPImage, timestampMs: Int) throws {
        // Async: automatically drops incoming camera frames when busy so results stay current
        try objectDetector?.detectAsync(image: mpImage, timestampInMilliseconds: timestampMs)
    }
}

extension ObjectDetectorStreamService: ObjectDetectorLiveStreamDelegate {
    func objectDetector(
        _ objectDetector: ObjectDetector,
        didFinishDetection result: ObjectDetectorResult?,
        timestampInMilliseconds: Int,
        error: (any Error)?
    ) {
        guard let result else { return }
        DispatchQueue.main.async {
            let detections = result.detections
        }
    }
}

Python

pippip install mediapipeSample code·Setup guide

Image

IMAGEStateless · single frameSynchronous call on independent still images. Resets temporal tracking state on every call.
import mediapipe as mp

BaseOptions = mp.tasks.BaseOptions
ObjectDetector = mp.tasks.vision.ObjectDetector
ObjectDetectorOptions = mp.tasks.vision.ObjectDetectorOptions
VisionRunningMode = mp.tasks.vision.RunningMode

options = ObjectDetectorOptions(
    base_options=BaseOptions(model_asset_path="efficientdet_lite0.tflite"),
    running_mode=VisionRunningMode.IMAGE,
    score_threshold=0.5,
    max_results=5,
)

with ObjectDetector.create_from_options(options) as detector:
    # Synchronous single-image call (stateless across images)
    mp_image = mp.Image.create_from_file("input.jpg")
    result = detector.detect(mp_image)

Video

VIDEOEvery frame · no dropsSynchronous tracking across decoded video frames with monotonic timestamps. Processes every frame in order and never drops frames.
import mediapipe as mp

BaseOptions = mp.tasks.BaseOptions
ObjectDetector = mp.tasks.vision.ObjectDetector
ObjectDetectorOptions = mp.tasks.vision.ObjectDetectorOptions
VisionRunningMode = mp.tasks.vision.RunningMode

options = ObjectDetectorOptions(
    base_options=BaseOptions(model_asset_path="efficientdet_lite0.tflite"),
    running_mode=VisionRunningMode.VIDEO,
    score_threshold=0.5,
    max_results=5,
)

with ObjectDetector.create_from_options(options) as detector:
    # Synchronous: processes every frame in order with monotonic timestamps (no frame drops)
    for mp_image, timestamp_ms in decoded_video_frames:
        result = detector.detect_for_video(mp_image, timestamp_ms)

Livestream

LIVE_STREAMAsync · drops frames when busyAsynchronous camera pipeline with a result callback. Automatically drops incoming frames when busy so results stay current with low latency.
import mediapipe as mp

BaseOptions = mp.tasks.BaseOptions
ObjectDetector = mp.tasks.vision.ObjectDetector
ObjectDetectorOptions = mp.tasks.vision.ObjectDetectorOptions
VisionRunningMode = mp.tasks.vision.RunningMode

def on_result(result, output_image: mp.Image, timestamp_ms: int):
    print(f"Detected {len(result.detections)} object(s) at {timestamp_ms} ms")

options = ObjectDetectorOptions(
    base_options=BaseOptions(model_asset_path="efficientdet_lite0.tflite"),
    running_mode=VisionRunningMode.LIVE_STREAM,
    score_threshold=0.5,
    max_results=5,
    result_callback=on_result,
)

with ObjectDetector.create_from_options(options) as detector:
    # Async: automatically drops incoming camera frames when busy so results stay current
    mp_image = mp.Image(image_format=mp.ImageFormat.SRGB, data=rgb_frame)
    detector.detect_async(mp_image, timestamp_ms)

Task details

This section describes the capabilities, inputs, outputs, and configuration options of this task.

Features

  • Input image processing - Processing includes image rotation, resizing, normalization, and color space conversion.
  • Label map locale - Set the language used for display names.
  • Score threshold - Filter results based on prediction scores.
  • Top-k detection - Limit the number of detection results.
  • Label allowlist and denylist - Specify the categories detected.
Inputs
The Object Detector accepts an input of one of the following data types:
  • Still images
  • Decoded video frames
  • Live video feed

See Prepare vision input data for MPImage, mp.Image, and browser media examples.

Outputs
The Object Detector outputs the following results for detected objects:
  • Category of object
  • Probability score
  • Bounding box coordinates

Configuration options

This task has the following configuration options:

Web

Option Name Description Value Range Default Value
runningMode Sets the running mode for the task:

IMAGE: Single image inputs.
VIDEO: Decoded video frames and live camera streams.		 {IMAGE, VIDEO} IMAGE
displayNamesLocale Locale code for display names provided in the model metadata, if available. You can add localized labels to custom models with the LiteRT Metadata Writer API. Locale code "en"
maxResults Maximum number of top-scored detection results to return (-1 returns all results). Integer -1
scoreThreshold Prediction score threshold overriding the model metadata default. Detections lower than this score are rejected. Float [0.0, 1.0] Not set
categoryAllowlist Optional list of allowed category names. Mutually exclusive with categoryDenylist. List of strings Not set
categoryDenylist Optional list of excluded category names. Mutually exclusive with categoryAllowlist. List of strings Not set

Android

Option Name Description Value Range Default Value
runningMode Sets the running mode for the task:

IMAGE: Single image inputs.
VIDEO: Decoded frames of a video.
LIVE_STREAM: Livestream of camera input data (requires asynchronous listener/delegate).		 {IMAGE, VIDEO, LIVE_STREAM} IMAGE
displayNamesLocale Locale code for display names provided in the model metadata, if available. You can add localized labels to custom models with the LiteRT Metadata Writer API. Locale code "en"
maxResults Maximum number of top-scored detection results to return (-1 returns all results). Integer -1
scoreThreshold Prediction score threshold overriding the model metadata default. Detections lower than this score are rejected. Float [0.0, 1.0] Not set
categoryAllowlist Optional list of allowed category names. Mutually exclusive with categoryDenylist. List of strings Not set
categoryDenylist Optional list of excluded category names. Mutually exclusive with categoryAllowlist. List of strings Not set
resultListener Asynchronous callback for LIVE_STREAM mode results. OutputHandler.ResultListener<ObjectDetectorResult, MPImage> Not set
errorListener Optional error listener for LIVE_STREAM mode. ErrorListener Not set

iOS

Option Name Description Value Range Default Value
runningMode Sets the running mode for the task:

IMAGE: Single image inputs.
VIDEO: Decoded frames of a video.
LIVE_STREAM: Livestream of camera input data (requires asynchronous listener/delegate).		 {IMAGE, VIDEO, LIVE_STREAM} IMAGE
displayNamesLocale Locale code for display names provided in the model metadata, if available. You can add localized labels to custom models with the LiteRT Metadata Writer API. Locale code "en"
maxResults Maximum number of top-scored detection results to return (-1 returns all results). Integer -1
scoreThreshold Prediction score threshold overriding the model metadata default. Detections lower than this score are rejected. Float [0.0, 1.0] Not set
categoryAllowlist Optional list of allowed category names. Mutually exclusive with categoryDenylist. List of strings Not set
categoryDenylist Optional list of excluded category names. Mutually exclusive with categoryAllowlist. List of strings Not set
objectDetectorLiveStreamDelegate Asynchronous delegate for LIVE_STREAM mode results (invoked on a serial background queue). ObjectDetectorLiveStreamDelegate Not set

Python

Option Name Description Value Range Default Value
running_mode Sets the running mode for the task:

IMAGE: Single image inputs.
VIDEO: Decoded frames of a video.
LIVE_STREAM: Livestream of camera input data (requires result_callback).		 {IMAGE, VIDEO, LIVE_STREAM} IMAGE
display_names_locale Locale code for display names provided in the model metadata, if available. You can add localized labels to custom models with the LiteRT Metadata Writer API. Locale code "en"
max_results Maximum number of top-scored detection results to return (-1 returns all results). Integer -1
score_threshold Prediction score threshold overriding the model metadata default. Detections lower than this score are rejected. Float [0.0, 1.0] Not set
category_allowlist Optional list of allowed category names. Mutually exclusive with category_denylist. List of strings Not set
category_denylist Optional list of excluded category names. Mutually exclusive with category_allowlist. List of strings Not set
result_callback Asynchronous callback function for LIVE_STREAM mode results. Callable Not set

Task output

Each ObjectDetectorResult contains a detections list of bounding boxes and ranked class labels:

Field Type / Coordinate Space Description
boundingBox BoundingBox (pixels) Unnormalized pixel rectangle (originX, originY, width, height) in input image coordinates (bounding_box in Python).
categories Category[] (score in [0.0, 1.0]) Ranked object class labels (categoryName, displayName, index, score).

The following example shows two detected objects (dog) with their pixel bounding boxes and confidence scores:

{
  detections: [
    {
      boundingBox: { originX: 355, originY: 133, width: 190, height: 206 }, // pixel coordinates
      categories: [{ index: 17, score: 0.7383, categoryName: "dog" }]
    },
    {
      boundingBox: { originX: 103, originY: 15, width: 138, height: 369 },
      categories: [{ index: 17, score: 0.7305, categoryName: "dog" }]
    }
  ]
}

Models

Object Detector pipeline overview

The Object Detector requires an object detection model to be downloaded and stored in your project directory. If you don't already have a model, start with the default, recommended model. The other models presented in this section make trade-offs between latency and accuracy.

The EfficientDet-Lite0 model uses an EfficientNet-Lite0 backbone with a 320x320 input size and BiFPN feature network. The model was trained with the COCO dataset, a large-scale object detection dataset that contains 1.5 million object instances and 80 object labels. See the full list of supported labels. EfficientDet-Lite0 is available as an int8, float16, or float32 model. This model is recommended because it strikes a balance between latency and accuracy. It is both accurate and lightweight enough for many use cases.

Model name Input shape Quantization type Versions
EfficientDet-Lite0 (int8) 320 x 320 int8 Latest
EfficientDet-Lite0 (float 16) 320 x 320 float 16 Latest
EfficientDet-Lite0 (float 32) 320 x 320 None (float32) Latest

EfficientDet-Lite2 model

The EfficientDet-Lite2 model uses an EfficientNet-Lite2 backbone with a 448x448 input size and BiFPN feature network. The model was trained with the COCO dataset, a large-scale object detection dataset that contains 1.5 million object instances and 80 object labels. See the full list of supported labels. EfficientDet-Lite2 is available as an int8, float16, or float32 model. This model is generally more accurate than EfficientDet-Lite0, but is also slower and more memory intensive. This model is appropriate for use cases where accuracy is a higher priority than speed and size.

Model name Input shape Quantization type Versions
EfficientDet-Lite2 (int8) 448 x 448 int8 Latest
EfficientDet-Lite2 (float 16) 448 x 448 float 16 Latest
EfficientDet-Lite2 (float 32) 448 x 448 None (float32) Latest

SSD MobileNetV2 model

The SSD MobileNetV2 model uses a MobileNetV2 backbone with a 256x256 input size and SSD feature network. The model was trained with the COCO dataset, a large-scale object detection dataset that contains 1.5 million object instances and 80 object labels. See the full list of supported labels. SSD MobileNetV2 is available as a float 16 and float 32 model. This model is faster and lighter than EfficientDet-Lite0, but is also generally less accurate. This model is appropriate for use cases that require a fast, lightweight model that sacrifices some accuracy.

Model name Input shape Quantization type Versions
SSDMobileNet-V2 (float 16) 256 x 256 float 16 Latest
SSDMobileNet-V2 (float 32) 256 x 256 None (float32) Latest

Model requirements and metadata

This section describes the requirements for custom models if you decide to build a model to use with this task. Custom models must be in LiteRT format and must include metadata describing the operating parameters of the model.

Design requirements

Input Shape Description
Input image Float32 tensor of shape[1, height, width, 3] The normalized input image.
Output Shape Description
detection_boxes Float32 tensor of shape [1, num_boxes, 4] Box location of each detected object.
detection_classes Float32 tensor of shape [1, num_boxes] Indexes of the class names for each detected object.
detection_scores float32 tensor of shape [1, num_boxes] Prediction scores for each detected object.
num_boxes Float32 tensor of size 1 The number of detected boxes.

Metadata requirements

Parameter Description
input_norm_mean The mean value used in the input tensor normalization.
input_norm_std The standard deviation used in the input tensor normalization.
label_file_paths The paths to the category tensor label files. If the model does not have any label files, pass an empty list.
score_calibration_md Information on the score calibration operation in the classification tensor. This parameter is not required if the model does not use score calibration.

Task benchmarks

Here are the task benchmarks for the preceding pre-trained models. The latency result is the average latency on Pixel 6 using CPU / GPU.

Model name Device CPU latency GPU latency
EfficientDet-Lite0 float32 model Pixel 6 61.30ms 27.83ms
EfficientDet-Lite0 float16 model Pixel 6 53.97ms 27.97ms
EfficientDet-Lite0 int8 model Pixel 6 29.31ms -
EfficientDet-Lite2 float32 model Pixel 6 197.98ms 41.15ms
EfficientDet-Lite2 float16 model Pixel 6 198.77ms 47.31ms
EfficientDet-Lite2 int8 model Pixel 6 70.91ms -
SSD MobileNetV2 float32 model Pixel 6 36.30ms 24.01ms
SSD MobileNetV2 float16 model Pixel 6 37.35ms 28.16ms