Self-learning object recognition system that labels new objects in images and video using ConvNeXt feature embeddings, cosine similarity, and dynamic thresholding - inspired by work at Skylark Labs on self-teaching vision pipelines.
- Cosine similarity scoring - Match new objects against a growing feature library
- Dynamic label assignment - Manual and automated thresholding for new classes
- ConvNeXt embeddings - Strong visual features for fine-grained recognition
- Image + video pipelines - Batch image processing and real-time video labeling
| Layer | Tools |
|---|---|
| Deep learning | PyTorch, ConvNeXt, torchvision |
| Vision | OpenCV, PIL |
| Math | NumPy, SciPy |
| Data | DataLoader, custom transforms |
Reference images → ConvNeXt features → Similarity bank
New frame/image → ConvNeXt features → Cosine match → Label assignment
- Incremental object learning without full retraining
- Quality control and defect recognition prototypes
- Research demos for continual / self-supervised vision
Developed alongside Skylark Labs internship work on YOLOv8 + DreamSim + BoT-SORT tracking pipelines.