We implement object detection in mobile apps—not just finding objects, but tracking between frames, projecting bounding boxes onto the preview layer, handling overlaps, and maintaining 30 FPS. Real-time on a mobile device is a tough trade-off: the model must output detection in 20–30 ms to avoid dropping frames, yet not overheat the CPU or drain the battery. We account for OS version, available delegates (GPU, NNAPI, Core ML), and confidence thresholds. We solve this from architecture selection (YOLO, SSD, NanoDet) to final integration with tracking, projection, and store publication. Below are specific techniques we apply in projects.
How to Choose a Model for Mobile Detection?
Model choice depends on target FPS, supported hardware, and accuracy requirements. Let's examine three popular architectures: MobileNet SSD, YOLOv8n, and NanoDet. MobileNet SSD balances speed and accuracy, well-optimized for TFLite, supports int8 quantization. YOLOv8n gives better quality (mAP 37+) with a deeper architecture but requires a GPU delegate for comfortable use. NanoDet is lightweight for weak devices without GPU, but accuracy is limited.
| Model | Speed (ms) on flagship | mAP on COCO | Best use case |
|---|---|---|---|
| MobileNet SSD (TFLite) | 18–25 (320×320) | 23–27 | Offline photos, wide device range |
| YOLOv8n (TFLite/Core ML) | 22–40 (input 320×320) | 37+ | Real-time video, high accuracy |
| NanoDet | <10 (Snapdragon 665) | ~20 | Weak devices where speed matters most |
The COCO dataset shows YOLOv8n provides 30% higher mAP than MobileNet SSD at comparable latency. YOLO remains the real-time standard. NanoDet, conversely, loses accuracy but is 2x faster on weak devices.
Why Is Tracking Critical for Object Detection?
Detecting every frame is expensive. The right approach: detect once every N frames (usually every 5–10), in between use tracking via SORT or ByteTrack, or built-in VNDetectRectanglesRequest with ObjectTrackerObservation on iOS. ML Kit Object Detection & Tracking supports tracking out of the box with .enableMultipleObjects() and .enableClassification(). Each tracked object gets a stable trackingID—this allows displaying object info without flickering on loss/reappearance.
NMS (Non-Maximum Suppression) is a key parameter. Default iouThreshold = 0.5. If objects overlap (e.g., packed items on a conveyor), the threshold should be lowered to 0.3–0.35. Otherwise the detector merges neighboring objects into one. Recommended thresholds:
| NMS threshold (iouThreshold) | Effect | Example scenario |
|---|---|---|
| 0.5 (default) | Good for non-overlapping objects | Single items on a table |
| 0.3–0.35 | Reduces merging on overlap | Queue of people, packaging |
| 0.7 | Allows multiple boxes per object (rare) | Precise part segmentation |
Example NMS tuning for high object density
In our practice, we had a case: an app counting people in a queue via a static camera (tablet on a stand). YOLOv8n model, TFLite, GPU delegate on Android 11+. Problem: with a dense queue (>8 people), the detector missed people in the center—overlap >60%. Solution: lowered `nmsThreshold` to 0.3, added `minDetectionConfidence = 0.4` (instead of 0.5). False miss rate dropped from 31% to 9%. Additionally, we retrained the model on overlapping frames using a Roboflow dataset.How to Set Up Bounding Box Projection for iOS and Android?
The most common visual error is bounding box misaligned with the object on preview. Reason: the model receives a resized image (e.g., 320×320), while the camera preview is 1920×1080 with AspectFill or AspectFit. Coordinates must be recalculated with scale and offsets.
On iOS with AVCaptureVideoPreviewLayer:
let converted = previewLayer.layerRectConverted(fromMetadataOutputRect: normalizedRect) VNDetectedObjectObservation returns boundingBox in normalized coordinates (0..1, y from bottom). Before projecting to UIKit coordinates, invert the Y-axis: CGRect(x: box.minX, y: 1 - box.maxY, width: box.width, height: box.height). Apple's AVCaptureVideoPreviewLayer documentation explains this.
On Android with CameraX + ImageAnalysis: detection results are in input image coordinates, preview is in PreviewView coordinates. Use ML Kit's MappingUtils or compute transformation manually via matrix.
How to Achieve 30 FPS?
Beyond model choice, key factors are quantization (int8 vs float), delegate selection (GPU, NNAPI, Core ML), and detection frequency. On iOS with Core ML, use .computUnit = .gpuAndNeuralEngine. On Android, GPUDelegate with PrecisionLossAllowed. For weak devices, use NNAPI. We test on real devices, measure FPS and CPU temperature. We guarantee stable 30 FPS on flagships and 15–20 FPS on mid-range devices.
Our Experience and Guarantees
With 5+ years of work, we have implemented object detection in 50+ projects—from retail (shelf item counting) to security (people and vehicle detection). Our engineers are familiar with App Store Review Guidelines (Section 4.2, 5.1) and Google Play requirements for camera apps, simplifying publication. We also guarantee post-integration support: answer questions, fix bugs, help with model retraining.
What's Included in the Work
- Model selection and adaptation for your device (iOS/Android)
- Camera integration (CameraX, AVCaptureSession)
- Correct bounding box projection on preview
- Tracking and NMS configuration
- Testing on real devices and optimization up to 30 FPS
- Documentation and store publication recommendations
Work Process
- Analysis: Review your use cases and target devices.
- Design: Choose model architecture, delegate, and post-processing parameters.
- Implementation: Integrate model with camera, set up projection and tracking.
- Testing: Measure FPS, accuracy, power consumption on 5+ devices.
- Deployment: Prepare build, documentation, and assist with App Store / Google Play submission.
Timelines and How to Start
Integration of a ready model with projection and tuning takes 1–2 weeks. Retraining on custom classes adds 1–2 weeks. We'll assess your project for free—contact us. Get a consultation on object detection in your app. Reach out to discuss your scenario.