object_detection 0.4.0

On-device object detection using MediaPipe EfficientDet-Lite models in LiteRT (formerly TensorFlow Lite) format. 80 COCO classes, fully offline.

0.4.0

• Default precision is now Precision.fp32 instead of fp16. This changes numeric output. flutter_litert 3.8.0 changed its own default for the same reason: across 29 published detection models measured on five GPUs, fp16 matched a plain-CPU reference for only about a fifth of them, while fp32 matched every model that compiled. These graphs emit pixel-space coordinates and landmark positions, and fp16 carries about three decimal digits of mantissa, so the error lands directly on output geometry. The cost is real and worth stating plainly: fp32 is a median 29.9% slower on GPU across those five GPUs, with Apple M4 the lone exception at 6.5% faster. Pass precision: Precision.fp16 explicitly to restore the previous behaviour, ideally per model and validated on your target GPU.
• Pin flutter_litert to ^3.8.0.
• Add the LiteRT Next CompiledModel engine as an opt-in alternative to the Interpreter path, matching face_detection_tflite, pose_detection and hand_detection. ObjectDetector.create / initialize and ObjectDetection.createCompiledFromBuffer take useCompiledModel, accelerators and precision. Default remains the Interpreter path. Measured end-to-end on macOS (Apple Silicon): Lite0 1.3-2.1x faster, Lite2 2.8-3.3x faster.
• The compiled engine uses TensorBufferMode.hostMemory and the writeInput / dispatch / readOutput zero-copy path, so the detection heads (~1.8M floats per frame for Lite0, ~3.4M for Lite2) are decoded as views of model-owned memory instead of being copied into fresh Dart lists. Measured 1.4x faster than the managed-buffer runAsync path. Falls back to managed buffers where host memory is unavailable.
• Preprocessing now uses OpenCV's SIMD cvtColor + convertTo kernels rather than a per-pixel Dart loop, via the new bgrMatToSignedFloat32. Measured 4.1-4.5x faster at the tensor-conversion step, numerically equivalent to the scalar path (max absolute difference of one float32 ULP).
• convertImageToTensor skips cv.resize when the source already matches the target geometry and skips cv.copyMakeBorder when the letterbox is empty.
• The detection isolate reuses one input tensor buffer for its whole life instead of allocating 1.2 MB (Lite0) or 2.4 MB (Lite2) per frame. ObjectDetection.newInputBuffer() exposes the same for direct API users.
• Anchors are generated into a flat Float32List instead of one List<double> per anchor (19 206 of them for Lite0, 37 629 for Lite2), and the decode loop seeds its argmax at the score threshold and writes survivors into reusable typed scratch buffers. Measured 1.2x faster at the decode step, with identical output. New generateEfficientDetAnchorsFlat; generateEfficientDetAnchors is unchanged and now delegates to it.
• Add ObjectDetection.usesCompiledModel and ObjectDetection.activeAccelerators so callers can see which engine and accelerators a model actually compiled to.
• Re-export Accelerator and Precision from flutter_litert.
• Add a benchmark suite to the example: end-to-end and per-stage timings, invoke-vs-decode-vs-NMS attribution, a delegate/engine sweep, and an engine A/B that asserts the two engines agree. Each emits BENCH_JSON.
• The example app now defaults to the CompiledModel engine and carries a CM / Interpreter badge on all three screens that switches engines live, matching the face, pose and hand demos. Detector creation falls back to the Interpreter engine if CompiledModel cannot be created on the device, and the badge reports the engine actually in use. The package default is unchanged, so adding object_detection to an app never silently switches its inference engine.
• Warm the detector with one throwaway inference on creation. Without it the first timing shown after an engine switch is a cold CompiledModel number (Metal shader compilation) measured against a warm Interpreter one, which made the faster engine read as several times slower. Observed on Lite2: 140 ms cold vs 39 ms warmed, against 41 ms for the interpreter.
• The badge names the engine (CM / Interpreter) rather than a delegate. The other demos label this axis CM / XNN, but XNNPACK is only the interpreter path's delegate on desktop and Android; on iOS it is Metal, so an XNN label is wrong there.

0.3.0

• Fix confidence decoding for both bundled EfficientDet-Lite variants. Their class tensors are already produced by a TFLite LOGISTIC op; scores and thresholds are now used directly instead of applying sigmoid a second time. This removes widespread false positives and restores calibrated confidence values.
• Bump ObjectDetector.modelVersion because postprocessing output changes for the same input bytes.
• Add real-model regressions for Lite0 and Lite2 blank/uniform inputs.
• Breaking: trim the flutter_litert convenience re-export surface. Removed 16 symbols that this package never used and never documented: createNHWCTensor4D, fillNHWC4D, allocTensorShape, flattenDynamicTensor, sigmoid, sigmoidClipped, bgrBytesToRgbFloat32, packYuv420, YuvPlane, YuvLayout, PackedYuv, CameraPlane, coverFitScaleOffset, drawLandmarkMarker, drawSkeletonConnections, drawBoundingBoxOutline. If you relied on any of these, import them directly from package:flutter_litert/flutter_litert.dart.
• Deprecate OutputTensorInfo, collectOutputTensorInfo and testCollectOutputTensorInfo. They are byte-identical to copies in face_detection_tflite and are not called anywhere in this package outside their own test hook; flutter_litert 3.6.0 provides collectOutputShapes, which returns shapes without materializing tensor buffers. Deprecated rather than removed, since they are public via part.
• Update flutter_litert -> 3.6.0.
• Expand the README live camera section with the full production pipeline (frame throttling, orientation handling, cover-fit overlay mapping).

0.2.3

• Update flutter_litert -> 3.5.0

0.2.2

• Update flutter_litert -> 3.4.1 (web CompiledModel WebGPU compile watchdog: a compile attempt that never settles now falls back to WASM instead of hanging). No API change.

0.2.1

• Update flutter_litert -> 3.3.1

0.2.0

• Update flutter_litert -> 3.2.0
• Import native-only flutter_litert APIs via package:flutter_litert/native.dart so they resolve under static analysis (flutter_litert 3.2.0 moved InterpreterFactory, IsolateWorkerBase, and TensorFloat32Views behind the native conditional export). No runtime or API change.

0.1.2

• Update flutter_litert -> 3.1.1

0.1.1

• Update flutter_litert -> 2.8.3

0.1.0

• Update flutter_litert -> 2.8.0
• Complete Swift Package Manager migration: example apps build via SPM without CocoaPods

0.0.8

• Remove unused Darwin podspecs for Dart-only iOS/macOS plugin registration.

0.0.7

• Update flutter_litert -> 2.5.8

0.0.6

• Update flutter_litert -> 2.5.5

0.0.5

• Update flutter_litert to 2.5.3

0.0.4

• Update flutter_litert -> 2.5.2

0.0.3

• Update flutter_litert -> 2.5.0

0.0.2

• Update flutter_litert -> 2.4.1

0.0.1 - 2026-04-27

• Initial release.
• On-device object detection over 80 COCO classes.
• Two model variants: EfficientDet-Lite0 (default, 320×320) and EfficientDet-Lite2 (448×448).
• Per-call options: scoreThreshold, maxResults, categoryAllowlist, categoryDenylist.
• Background isolate for inference; UI thread is never blocked.
• Image input variants: encoded bytes, file path, cv.Mat, raw pixel bytes, CameraImage, CameraFrame.
• Hardware-accelerated by default: Metal on iOS, XNNPACK elsewhere.
• Cross-platform: Android, iOS, macOS, Windows, Linux.