This project explores building a custom object detection pipeline for basic sign language gestures using the TensorFlow Object Detection API (TFOD) and an SSD-based architecture.
The focus of this repository is on data preparation, annotation, TFRecord generation, and pipeline configuration, rather than final model training.
- Custom dataset of hand gesture images representing basic sign language classes
- Images annotated in Pascal VOC (XML) format
- Conversion of annotations to TFRecords
- Configuration of an SSD MobileNet pipeline using TFOD
The project was initially intended for end-to-end training; however, due to recent TensorFlow API deprecations and managed notebook environment constraints, training with TFOD is no longer reliably supported on platforms like Google Colab or Kaggle without custom Docker setups.
- Designed a complete object detection dataset pipeline
- Created and validated:
label_map.pbtxttrain.record/test.record(not included in repo)- SSD
pipeline.config
- Debugged real-world dependency and compatibility issues across:
- TensorFlow versions
- Python versions
- Cloud notebook environments
- Gained hands-on experience with production-style ML project structure
The following are intentionally not included:
- Training images and XML annotations
- TFRecord files
- Model checkpoints and exported models
This keeps the repository lightweight and avoids sharing large or private datasets.
- Python
- TensorFlow Object Detection API
- SSD MobileNet
- LabelImg
- TFRecords
- XML (Pascal VOC)
While TensorFlow Object Detection API is now considered a legacy stack, this project served as a strong foundation for understanding object detection workflows and has informed future work using modern frameworks such as PyTorch and YOLO-based models.
Rahul