Where's My Stuff? - A Smart Inventory App

Overview

"Where's My Stuff?" is a mobile application designed to help you keep track of your belongings. Using your phone's camera, you can quickly scan items, and the app will automatically identify them using a pruned MobileNet V4 Small model. You can then assign your items to a specific place, making it easy to remember where you've stored them.

The app features a clean, modern interface with both light and dark themes, and it's built with simplicity and ease of use in mind.

This app was built from model pruning and finetuning to flutter app development in 3 days as a challenge in fast prototyping. AI agents were used to assist with handling of the coding parts for speed.

Features

• AI-Powered Object Recognition: The app uses a pre-trained TensorFlow Lite model to automatically identify and suggest labels for your items with a confidence score of 25% or higher.
• Camera-Based Scanning: Simply point your camera at an item and take a picture. The app handles the rest.
• Multi-Label Selection: After scanning, you can select multiple accurate labels from a list of suggestions.
• Place-Based Organization: Assign your scanned items to a "place" (e.g., "Garage," "Kitchen Cupboard") to keep your inventory organized.
• Persistent Inventory: Your inventory is saved locally on your device, so you can access it anytime.
• Theme Customization: Choose between a light or dark theme, and the app will remember your preference.
• Modern, Responsive UI: The app is built with Flutter and Material 3, providing a beautiful and intuitive user experience on any device.

Technical Details

• Framework: Flutter 3.x
• Language: Dart 3.x
• State Management: flutter_riverpod for efficient and scalable state management.
• Database: sqlite3 and sqlite3_flutter_libs for reliable local data storage.
• Machine Learning: tflite_flutter for running the on-device TensorFlow Lite model.
• Camera: camera and camera_android_camerax for seamless camera integration.
• UI: Material 3 design principles for a modern and consistent look and feel.

How to Build and Run

Download APK and install on device directly.

Alternatively:

1. Clone the repository:

   git clone <https://github.com/Bhabesh-Rath/Where-s-my-stuff-.git>

2. Get dependencies:

   flutter pub get

3. Run the app:

   flutter run

Screenshots

Classification popup after clicking a photo	List created and saved	Search function

Design Decisions & Constraints

• Problem framing: Most household inventory apps require cloud sync or manual text entry. The goal was a camera-first, fully offline app that works on mid-to-low tier Android hardware without internet dependency.

• Model selection: MobileNet V4 Small was chosen over larger classification models because of its favourable accuracy and low latency. Pruning reduced model size further to meet the <50ms inference target on device.

• Structured Pruning over Unstructured: Structure pruning removes entire filters or channels, making the hardware natively faster on standard hardware without requirinng sparce inference engines. It affects inference speed making it the better choice for a low latency target compared to unstructured which might lead to better sparcity but has less optimized memory access and might cause latency issues.

• COCO Dataset for finetuning: It is a large dataset with 80 object categories which is a good starting point for a model that has to classify objects present in an image. It was prioritized as it has enough household objects to verify usability without requiring the curation of a custom dataset in a 3 day constraint.

• Why TFLite over ONNX: TFLite's Android integration with flutter_tflite is better documented and more stable for production APKs. ONNX Runtime Mobile was evaluated but deprioritised due to integration complexity within the 3-day constraint.

• Confidence threshold of 25%: Set deliberately low because the use case (personal item tracking) benefits from over-suggestion — users can dismiss irrelevant labels. A higher threshold would increase silent misses, which is the worse failure mode here.

• Local-first storage: SQLite chosen over cloud sync to keep the app functional without internet and avoid data privacy concerns for personal inventory data.

Future Improvements

• UI Refinement: Label display optimizatiion on scan label popups; duplicate place-name validation.
• Better Classification: Finetuning with a different Dataset or using a different base model for more accurate labeling.
• Object Marking: A segmentation model to separate out different objects to aid with classification.
• Custom Labels and continuous finetuning: Allow users to add their own custom labels for items that the AI model may not recognize and then train the model during downtime to better it's classifications with user data.