Qualcomm VisionX Project Name : DetectX Subtitle : "Round 1 Submission for IIT Bombay Techfest 2024" Team Details: Team Name : ETHAN Team Members : ANIRUTH PRABU SRIVATSAN K RIJUL NAMBIAR P Team ID : Qual-230345 Email: [email protected] Phone: 8838818224 Home About Content Result
Introduction Objective: Automate object detection in images and videos using TensorFlow Object Detection API. Perform instance segmentation to identify specific object areas. Key Features: Real-time detection Pretrained models for accuracy and speed Applications: Autonomous vehicles, surveillance, retail analytics. Home About Content Result
Technology Stack Frameworks: TensorFlow Object Detection API, NumPy, PIL, and Matplotlib. Tools: Python, Jupyter Notebook, TensorFlow Model Zoo. Models Used: SSD MobileNet v1 for object detection. Mask R-CNN for instance segmentation. Environment: Python 3.8, TensorFlow 2.x. Home About Content Result
Dataset and Pretrained
Models Dataset: COCO Dataset (Common Objects in Context). Model Zoo: SSD MobileNet v1: Lightweight model for object detection. Mask R-CNN: Advanced model for instance segmentation. Home About Content Result
Workflow Overview Steps: a. Setup Environment: Install dependencies, clone TensorFlow models, compile protobufs. b. Load Pretrained Models: Use SSD MobileNet v1 and Mask R-CNN. c. Process Images/Videos: Convert images to tensors for inference. d. Run Inference: Detect objects and visualize results. e. Display Output: Annotated images/videos with bounding boxes and labels. Home About Content Result
Detailed Action Plan
Phase 1: Setup & Model Selection Phase 2: API Development & Integration Choose and fine-tune models for Implement FastAPI endpoints for AI Conversational AI (DistilBERT/OpenChatKit) interactions and code generation. and Knowledge Base (T5). Build frontend interface using ReactJS and Select appropriate code generation model Streamlit for rapid testing. (T5/Codex).
Phase 3: Testing & Optimization
Perform alpha and beta testing with real users for Conversational AI. Optimize knowledge base for query accuracy and compliance. Prototype code generation tool and gather developer feedback. Home About Content Result
Workflow Overview Inputs: Batch of RGB images (uint8 format). Outputs: Bounding boxes, detection scores, and class labels. Instance segmentation masks for Mask R- CNN. Inference Function: run_inference_for_single_image: Processes single images for detection. Outputs cleaned data for easy visualization. Home About Content Result
Results and Visualization
Object Detection Results:
Annotated images showing detected objects with bounding boxes and labels. Instance Segmentation Results: Segmented masks highlighting detected object areas. Home About Content Result
Experience Our Prototype in Google Drive: https://drive.google.com/drive/folders/1R6m--
BWw4cMkEwzGqYRBTfGMYj-8VlFZ?usp=sharing Home About Content Result
Experience Our Prototype in Youtube: https://youtu.be/mOKVw-0DdGw?si=SSbWjaBTgG_8qhcY
Home About Content Result
Conclusion and Future Work Conclusion: Successfully implemented object detection and instance segmentation. High accuracy and reliable visualization with pretrained models. Future Enhancements: Use custom datasets for specific applications. Deploy models on edge devices for real-time inference. Extend to multi-object tracking in videos.
