Object Classification in Mobile Apps
We integrate object classification into mobile apps turnkey. Unlike detection, the model outputs a single probability vector per class. It seems simpler, but this is where developers most often make mistakes: choosing the wrong confidence threshold, ignoring post-processing, and forgetting about UX. In this article, we break down how to avoid these pitfalls and make classification reliable and user-friendly. Experience with Core ML, ML Kit, EfficientNet—over 5 years in mobile development.
How to Choose a Classification Model?
For top-1000 classes (products, animals, household items), MobileNetV3-Large or EfficientNetB0/B1 work well. They run out-of-the-box via ML Kit Image Labeling or Core ML with models from the Apple Model Gallery. For a narrow domain (specific product types, manufacturing defects), a fine-tuned model is required.
Comparison of popular options:
| Model |
Size |
Accuracy (top-1) |
Applicability |
| MobileNetV3-Large |
~5 MB |
75.6% |
General, fast |
| EfficientNetB0 |
~8 MB |
77.1% |
Balance of speed and accuracy |
| Custom (fine-tuned) |
5-10 MB |
80-90% (dataset dependent) |
Narrow domain |
| Solution type |
When to choose |
Example |
| Pre-built model |
Top-1000 classes, quick start |
ML Kit Image Labeling |
| Fine-tuning |
10–50 classes, 200–500 examples per class |
Custom EfficientNet |
| Few-shot learning |
Fewer than 50 examples per class |
Prototypical Networks |
Fine-tuning on a custom dataset: take a pre-trained backbone (MobileNetV2, EfficientNetB0), freeze the lower layers, train only the top layers on your data. For 10–50 classes, 200–500 examples per class with proper augmentation suffice. Fewer—requires a few-shot approach (Prototypical Networks). After training, convert to .mlmodel (iOS) or .tflite (Android), adding metadata with class names and normalization parameters. EfficientNet is a good balance of speed and accuracy.
How to Choose the Confidence Threshold?
The most frequent UX mistake is showing classification results without considering confidence. The model always returns a probability distribution; argmax always picks a 'winner'—even when the model is unsure. If the top-1 class scores 0.23 with the next at 0.21, that's not classification—it's randomness.
The right approach: set a threshold (typically 0.5–0.7 depending on the task). If top-1 falls below the threshold, show 'unable to determine' or ask for a retake. For critical tasks (medical, legal documents)—also check the entropy of the distribution. According to Core ML documentation, the threshold is set via the confidenceThreshold parameter in VNCoreMLRequest.
On iOS via VNCoreMLRequest:
request.imageCropAndScaleOption = .centerCrop
let observations = results as? [VNClassificationObservation]
let confident = observations?.filter { $0.confidence > 0.65 }
On Android via ML Kit ImageLabeling:
val options = ImageLabelerOptions.Builder()
.setConfidenceThreshold(0.65f)
.build()
Why Is Post-Processing Important?
Showing top-3 classes with percentages is appropriate for educational and consumer apps. For business apps (automation, warehouse), you need one confident result or nothing. Post-processing includes not only the threshold but also checking the gap between top-1 and top-2: if the difference is small, ask for user confirmation.
Case study: an inventory app for a warehouse, classifying 87 SKUs via a custom EfficientNetB0 model. The initial threshold of 0.5 gave 12% false positives. After analyzing the confusion matrix, we found that 80% of errors occurred between SKUs with similar packaging. We added a second level: if top-2 and top-3 together exceed 0.4, the operator is prompted to confirm. False positives dropped to 2.1%.
UI for results: don't overwhelm users with numbers. A progress bar or color indicator (green/yellow/red) is perceived better than '73.4%'. Animating the result appearance via withAnimation (SwiftUI) or ObjectAnimator (Android) reduces the feeling of a 'cold' model response.
What Is Included
When you order object classification implementation, you get:
- Task analysis and model architecture selection
- Fine-tuning on your dataset (if needed)
- Conversion to
.mlmodel (iOS) / .tflite (Android)
- Integration into the app with data flow setup
- Confidence threshold tuning and post-processing
- UI development for result display
- Testing on real data and error correction
- Documentation and team support
Work Process
-
Analysis — understand the task, collect dataset, choose metrics
-
Design — determine architecture, processing pipeline
-
Implementation — fine-tuning, conversion, integration
-
Testing — unit tests, UI tests, A/B tests on confidence
-
Deployment — release to App Store / Google Play, monitoring
Timelines and Pricing
Integration of a ready-made model into an existing app: 3–5 days. Fine-tuning a custom model plus integration: 1–2 weeks. Pricing is determined individually—we evaluate your project after an initial consultation. We guarantee quality and deadlines, backed by over 5 years of mobile development experience.
Want to implement object classification? Contact us for a free consultation and preliminary estimate. Order classification integration into your app today.
Machine Learning in Mobile Apps: CoreML, TFLite, and On-Device Models
We distinguish two fundamentally different approaches: an app with on-device AI and an app that simply calls a cloud API. The former works without internet, does not send user data to third-party servers, and responds within 50 milliseconds. The latter depends on network latency and pricing plans. Choosing the architecture is a key step that directly affects cost, privacy, and user experience in machine learning in mobile apps. Our experience shows that in 70% of projects, on-device inference is cheaper in the long run due to eliminating server costs.
How to Choose Between CoreML and TFLite for On-Device Inference?
CoreML — Apple's native framework for running ML models on device. Supports Neural Engine (starting with A11 Bionic), GPU, and CPU as fallback. Models are converted to .mlmodel format via coremltools from PyTorch, ONNX, or TensorFlow. Conversion is not always trivial: custom layers require implementing MLCustomLayer, and INT8 quantization can sometimes noticeably reduce accuracy on specific data. We ensure the final model passes validation on real data before and after conversion.
TensorFlow Lite — cross-platform alternative for Android and Flutter. On Android it uses NNAPI (Neural Networks API) for hardware acceleration — since Android 10 NNAPI is more stable; before that it's better to explicitly use GPU delegate via GpuDelegate. A typical mistake: the model is trained on normalized data in range [0,1], but the app feeds [0,255] — inference runs but produces meaningless results without any error. We include an automatic input data validation module in the SDK.
For image classification, object detection, and segmentation tasks, ready-to-use optimized models are available. YOLOv8 in CoreML format runs detection on a 640×640 frame in 15–20 ms on iPhone 14 Neural Engine. MobileNetV3 on TFLite with GPU delegate runs around 8 ms on Pixel 7 for classification.
| Parameter |
CoreML |
TFLite |
| Platforms |
iOS, macOS, watchOS |
Android, iOS, Linux, embedded |
| Hardware acceleration |
Neural Engine, GPU, CPU |
NNAPI, GPU (OpenCL/OpenGL), CPU |
| Quantization support |
FP16, INT8 (with coremltools) |
FP16, INT8, dynamic range |
| Custom operations |
Via MLCustomLayer (Swift) |
Via delegates (Java/Kotlin) |
| Model bundle size |
~3–5 MB (MobileNetV2 quantized) |
~2–4 MB |
What If You Need Text Generation On-Device?
Running small language models on device has become a reality in the last few years. Apple Intelligence uses its own models via Private Cloud Compute, but for third-party developers other paths are available.
llama.cpp with Metal backend on iOS is a working approach for phi-3-mini (3.8B parameters, 4-bit quantization, ~2.3 GB). Inference: 15–25 tokens/second on iPhone 15 Pro. For integration in Swift, use the Swift Package llama.swift or a wrapper via C interface llama.h. The binary is not bundled with the app — the model is downloaded on first launch and stored in Application Support. Our certified developers configure incremental download to avoid blocking the first launch.
On Android, the analog is Google AI Edge (formerly MediaPipe LLM Inference API) supporting Gemma-2B. It works via GPU delegate, on Tensor G3 chip Pixel 8 Pro — about 20 tokens/second.
Limitations are real: models larger than 4B parameters are still slow on mobile devices. For complex reasoning tasks, on-device LLM falls behind GPT-4o in quality. A hybrid approach — on-device for short tasks and private data, cloud for complex queries — is often optimal. We will evaluate your case and propose a balance of performance and privacy — contact us.
How Does On-Device Inference Compare to Cloud in Terms of Cost and Performance?
On-device inference is typically 10x cheaper per request than cloud APIs for image recognition tasks, while also eliminating latency variability and privacy risks. The table below summarizes the trade-offs.
| Criteria |
On-Device Inference |
Cloud API |
| Latency |
<50ms |
200–500ms (including network) |
| Cost per 1M requests |
$0 (no server) |
$10–50 (AWS Rekognition, Google Vision) |
| Privacy |
Data stays on device |
Data sent to server |
| Offline |
Yes |
No |
| Scalability |
No server scaling issues |
Need to provision API capacity |
For an app with 100k MAU running 10 image recognitions per user per month, on-device inference can save up to $5,000 monthly compared to cloud API. Get a free consultation on your ML architecture today.
Integrating OpenAI API and Other Cloud Models
For scenarios where cloud inference is acceptable, integrating OpenAI, Anthropic, or Google Gemini is an HTTP client + streaming SSE. In Swift, AsyncThrowingStream is convenient for streaming responses. In Kotlin, use Flow.
Critically: API keys must never be stored in the app bundle. Even an obfuscated key can be extracted from the IPA in 10 minutes using strings or frida. Correct architecture: mobile app → your own backend → OpenAI API. The backend controls rate limiting, logs requests, and protects the key.
What Is Included in the Work (Deliverables)
- Trained and quantized model for the target device (documentation with metrics)
- SDK for integration (Swift/Kotlin/Flutter) with call examples
- Performance tests on 3–5 real devices
- Instructions for OTA model updates
- Support during App Store / Google Play moderation (compliance with Guidelines 4.2, 5.1)
- 2 weeks of technical support after release
Typical Project Pipeline
-
Task analysis — measure latency, privacy, size, supported devices.
-
Model prototyping — in Python, evaluate accuracy on target data.
-
Conversion and quantization — for CoreML/TFLite with validation.
-
Integration into the app — model wrapped in a service layer (easy to swap CoreML ↔ TFLite ↔ cloud).
-
Testing — on real devices, measure FPS, RAM, battery.
-
Deployment — via TestFlight / Firebase App Distribution, monitor metrics.
Timelines: integration of a ready CoreML/TFLite model — 1–2 weeks, development of a custom model with mobile optimization — from 6 weeks, on-device LLM chat with personalization — 4–8 weeks.
Why We Take on Complex Cases?
10+ years of experience in mobile development, 50+ implemented AI/ML solutions, guarantee of compatibility with current iOS and Android versions. All projects undergo code review and load testing. The cost includes preparation of moderation documentation and training of your team.
Contact us — we will help you choose the architecture and implement ML in your app turnkey. Order an audit of your existing solution — we will assess the potential for server cost savings free of charge. In some projects, savings can reach significant amounts per month.
