Vision Shift

• 

  Our Uploading Page

• 

  Mask of the parts which are different from initial image

• 

  Highlighted the difference in Final image and Initial Image

Inspiration

The idea for Vision Shift came from the growing need for intelligent visual monitoring systems that can automatically detect changes between two images without human supervision. Whether it’s monitoring environmental changes, construction progress, or security footage, identifying visual differences is time-consuming and error-prone when done manually. We wanted to leverage deep learning to create an automated, scalable solution that could detect even the smallest changes between two frames, inspired by real-world use cases like satellite change detection, medical image comparison, and surveillance analytics.

What it does

Vision Shift is a deep learning–powered visual difference engine that compares two images — a “before” and an “after” — and highlights the areas of change with precision. The system generates a change mask that pinpoints altered regions and calculates the exact percentage of change. Users can simply upload two images via a simple and intuitive Streamlit interface, and the model instantly processes them to display both the visual difference and the quantitative results.

How we built it

We built Vision Shift using Python and PyTorch, implementing a U-Net architecture with a ResNet34 encoder for semantic segmentation. The model was trained on paired image data inspired by the MOT20 dataset, focusing on change detection patterns. We used OpenCV and NumPy for image preprocessing and Matplotlib for visualization. The frontend was built with Streamlit, enabling easy interaction for non-technical users. The app runs inference through a lightweight API that takes uploaded images, processes them through the trained U-Net, and returns both the predicted mask and overlay result. Finally, we integrated version control using Git and GitHub, and designed it to be deployable on Streamlit Cloud or Hugging Face Spaces for seamless accessibility.

Challenges we ran into

One of the biggest challenges was handling dataset preparation, especially aligning paired images and generating ground truth masks. Training U-Net required high GPU resources, which led to performance and memory issues on local systems. We also faced difficulties in optimizing the model’s accuracy for subtle changes while avoiding false positives. Ensuring that the frontend interface remained lightweight and interactive while performing heavy backend computations was another key hurdle. Debugging file handling, checkpoint management, and balancing training performance with deployment size took considerable effort.

Accomplishments that we're proud of

We’re proud of successfully building a fully functional deep learning pipeline — from dataset preprocessing and model training to real-time inference via a user-friendly web interface. Vision Shift achieves smooth and accurate change detection with minimal setup and works even on standard CPU systems for testing. Another major achievement was designing a clear and clean frontend interface that anyone can use without coding knowledge. Completing this end-to-end AI project independently, from training to deployment readiness, was a major milestone.

What we learned

This project taught us how to integrate computer vision models with web frameworks, bridging the gap between machine learning research and usable applications. We gained hands-on experience with U-Net segmentation, PyTorch optimization, and Streamlit app development, as well as learned how to manage checkpoints, visualize masks, and optimize model performance. We also learned about practical issues like GPU utilization, dataset curation, and maintaining code modularity for scalability.

What's next for Vision Shift

In the future, we plan to enhance Vision Shift with multi-class change detection to categorize the types of changes, add real-time video feed analysis, and integrate cloud-based model hosting for faster inference. We aim to support more datasets and develop an API-based service for developers to integrate change detection into their own systems. Long-term goals include extending it to satellite imagery, medical imaging, and smart surveillance, making Vision Shift a universal, plug-and-play visual intelligence tool. just add markdowns wherever needed , nothing else should change

Built With

• flask
• numpy
• opencv
• python
• pytorch
• streamlit
• u-net