This project represents a cutting-edge approach in AI-driven object recognition and self-learning systems. The primary objective is to enhance the accuracy of object identification and labeling in dynamic environments. The system is designed to self-identify, self-label, and self-learn from the environment, becoming increasingly efficient over time.
The core of our method uses a detection model (YOLOv8) for accurately detecting objects within various frames. Once detected, DreamSim - a perceptual similarity model - analyzes and interprets the characteristics of each object. A specialized database of known objects is maintained, and a cosine similarity threshold determines when an object should be categorized as new. BoT-SORT tracks objects across frames and revises labels for accuracy.
- Significant improvement in the system's ability to accurately recognize and label objects, especially in dynamic scenes
- More efficient use and updating of the object database, leading to better recognition of new objects
- Decrease in incorrect labeling frequency, particularly when objects are partially visible or exiting the frame
- 40% boost in model accuracy through synthetic data generation
- 30% enhancement in detection accuracy for fine objects using super-resolution techniques
- Developed the self-labeling module - design and implementation of object recognition and labeling pipeline
- Integrated and optimized DreamSim for efficient feature extraction from detected objects
- Managed the database of known objects and fine-tuned cosine similarity thresholds for improved recognition
- Implemented BoT-SORT tracking for cross-frame object label revision and accuracy improvement
- Engineered synthetic data generation pipeline, increasing dataset diversity to drive 40% accuracy gain